Bereitstellen und Ausführen einer Vorlage für Elektronenstruktursimulation mit einem impliziten Lösungsmittelmodell
Diese in Zusammenarbeit mit der Cleveland Clinic entwickelte Vorlage besteht aus einem Workflow zur Berechnung der Grundzustandsenergie und freien Solvatationsenergie eines Moleküls in einem impliziten Lösungsmittel [1]. Diese Simulationen basieren auf der stichprobenbasierten Quantendiagonalisierungsmethode (SQD) [2-6] und dem Integralgleichungsformalismus des polarisierbaren Kontinuumsmodells (IEF-PCM) des Lösungsmittels [7].
Dieser Leitfaden verwendet die Vorlage mit einem Methanolmolekül als gelöstem Stoff, dessen Elektronenstruktur explizit simuliert wird, und Wasser als Lösungsmittel, das als kontinuierliches dielektrisches Medium approximiert wird. Um die Elektronenkorrelationseffekte in Methanol zu berücksichtigen und gleichzeitig das Gleichgewicht zwischen Rechenaufwand und Genauigkeit zu wahren, werden nur die -, - und einsamen Paarorbitale im aktiven Raum einbezogen, der mit SQD IEF-PCM simuliert wird. Diese Orbitalauswahl erfolgt mit der Atomic Valence Active Space (AVAS) Methode unter Verwendung der C[2s,2p], O[2s,2p] und H[1s] Atomorbitalkomponenten, was zu einem aktiven Raum von 14 Elektronen und 12 Orbitalen (14e,12o) führt. Die Referenzorbitale werden mit geschlossenerschaligem Hartree-Fock unter Verwendung des cc-pvdz-Basissatzes berechnet.
Workflow-Einführung
Dieser interaktive Leitfaden zeigt, wie du diese Funktionsvorlage auf Qiskit Serverless hochlädst und einen Beispiel-Workload ausführst. Die Vorlage ist als Qiskit-Pattern mit vier Schritten strukturiert:
1. Eingabe sammeln und Problem abbilden
Dieser Schritt nimmt die Geometrie des Moleküls, den ausgewählten aktiven Raum, das Solvatationsmodell, die LUCJ-Optionen und die SQD-Optionen als Eingabe. Er erzeugt dann die PySCF-Checkpoint-Datei, die die Hartree-Fock (HF) IEF-PCM-Daten enthält. Diese Daten werden im SQD-Teil des Workflows verwendet. Für den LUCJ-Teil des Workflows erzeugt der Eingabeabschnitt auch die Gasphasen-HF-Daten, die intern im PySCF-FCIDUMP-Format gespeichert werden.
Die Informationen aus der HF-Gasphasensimulation und die Definition des aktiven Raums werden als Eingabe verwendet. Wichtig ist, dass auch die benutzerdefinierten Informationen aus dem Eingabeabschnitt bezüglich der Fehlerunterdrückung, Anzahl der Shots, Optimierungsstufe des Schaltkreis-Transpilers und des Qubit-Layouts verwendet werden.
Es werden Ein-Elektronen- und Zwei-Elektronen-Integrale innerhalb des definierten aktiven Raums erzeugt. Die Integrale werden dann verwendet, um klassische CCSD-Berechnungen durchzuführen, die t2-Amplituden zurückgeben, die wir zur Parametrisierung des LUCJ-Schaltkreises verwenden.
2. Schaltkreis optimieren
Der LUCJ-Schaltkreis wird dann in einen ISA-Schaltkreis für die Ziel-Hardware transpiliert. Ein Sampler-Primitive wird dann mit einem Standardsatz von Fehlerminderungsoptionen instanziiert, um die Ausführung zu verwalten.
3. Schaltkreis ausführen
Die LUCJ-Berechnungen geben die Bitstrings für jede Messung zurück, wobei diese Bitstrings den Elektronenkonfigurationen des untersuchten Systems entsprechen. Die Bitstrings werden dann als Eingabe für die Nachbearbeitung verwendet.
4. Nachbearbeitung mit SQD
Dieser letzte Schritt nimmt die PySCF-Checkpoint-Datei mit den HF IEF-PCM-Informationen, die Bitstrings, die die von LUCJ vorhergesagten Elektronenkonfigurationen darstellen, und die im Eingabeabschnitt ausgewählten benutzerdefinierten SQD-Optionen als Eingabe. Als Ausgabe liefert er die SQD IEF-PCM-Gesamtenergie des niedrigsten Energie-Batches und die entsprechende freie Solvatationsenergie.
Optionen
Für diese Vorlage musst du Optionen zur Generierung des LUCJ-Schaltkreises und SQD-Laufzeitparameter angeben.
LUCJ-Optionen
Wenn der LUCJ-Quantenschaltkreis ausgeführt wird, wird eine Reihe von Stichproben erzeugt, die die rechnerischen Basiszustände aus der Wahrscheinlichkeitsverteilung des molekularen Systems darstellen. Um die Tiefe des LUCJ-Schaltkreises und seine Ausdrucksfähigkeit auszubalancieren, werden die Qubits, die den Spinorbitalen mit entgegengesetztem Spin entsprechen, mit Zwei-Qubit-Gates zwischen ihnen verbunden, wenn diese Qubits durch ein einzelnes Ancilla-Qubit benachbart sind. Um diesen Ansatz auf IBM-Hardware mit Heavy-Hex-Topologie zu implementieren, werden Qubits, die Spinorbitale mit gleichem Spin darstellen, durch eine Linientopologie verbunden, bei der jede Linie aufgrund der Heavy-Hex-Konnektivität der Ziel-Hardware eine Zickzack-Form annimmt, während die Qubits, die Spinorbitale mit entgegengesetztem Spin darstellen, nur bei jedem vierten Qubit eine Verbindung haben.
Klicke für weitere Details zu den erforderlichen Optionen:
Der Benutzer muss das initial_layout-Array bereitstellen, das den Qubits entspricht, die diesem Zickzack-Muster im lucj_options-Abschnitt der SQD IEF-PCM-Funktion entsprechen. Im Fall von SQD IEF-PCM (14e,12o)/cc-pvdz-Simulationen von Methanol haben wir das anfängliche Qubit-Layout gewählt, das der Hauptdiagonale der Eagle R3 QPU entspricht. Hier entsprechen die ersten 12 Elemente des initial_layout-Arrays [0, 14, 18, 19, 20, 33, 39, 40, 41, 53, 60, 61, ...] den Alpha-Spinorbitalen. Die letzten 12 Elemente [... 2, 3, 4, 15, 22, 23, 24, 34, 43, 44, 45, 54] entsprechen Beta-Spinorbitalen.
Wichtig ist, dass du die number_of_shots bestimmen musst, die der Anzahl der Messungen im LUCJ-Schaltkreis entspricht. Die Anzahl der Shots muss ausreichend groß sein, da der erste Schritt des S-CORE-Verfahrens auf den Stichproben im richtigen Teilchensektor beruht, um die anfängliche Approximation an die Grundzustands-Besetzungszahlverteilung zu erhalten.
Die Anzahl der Shots ist stark system- und hardwareabhängig, aber nicht-kovalente, fragmentbasierte und implizite Lösungsmittel SQD-Studien legen nahe, dass du die chemische Genauigkeit erreichen kannst, indem du diese Richtlinien befolgst:
- 20.000 - 200.000 Shots für Systeme mit weniger als 16 Molekülorbitalen (32 Spinorbitale)
- 200.000 Shots für Systeme mit 16 - 18 Molekülorbitalen
- 200.000 - 2.000.000 Shots für Systeme mit mehr als 18 Molekülorbitalen
Die erforderliche Anzahl von Shots wird durch die Anzahl der Spinorbitale im untersuchten System und durch die Größe des Hilbert-Raums beeinflusst, der dem ausgewählten aktiven Raum innerhalb des untersuchten Systems entspricht. Im Allgemeinen benötigen Instanzen mit kleineren Hilbert-Räumen weniger Shots. Weitere verfügbare LUCJ-Optionen sind Optimierungsstufe des Schaltkreis-Transpilers und Fehlerunterdrückungsoptionen. Beachte, dass diese Optionen auch die erforderliche Anzahl von Shots und die resultierende Genauigkeit beeinflussen.
SQD-Optionen
Wichtige Optionen in SQD-Simulationen umfassen sqd_iterations, number_of_batches und samples_per_batch. Im Allgemeinen kann eine geringere Anzahl von Stichproben pro Batch durch mehr Batches (number_of_batches) und mehr Iterationen von S-CORE (sqd_iterations) ausgeglichen werden. Mit mehr Batches können wir mehr Variationen der konfigurationellen Unterräume samplen. Da der Batch mit der niedrigsten Energie als Lösung für die Grundzustandsenergie des Systems genommen wird, können mehr Batches die Ergebnisse durch bessere Statistiken verbessern. Zusätzliche Iterationen von S-CORE ermöglichen es, mehr Konfigurationen aus der ursprünglichen LUCJ-Verteilung wiederherzustellen, wenn die Anzahl der Stichproben im korrekten Teilchensektor gering ist. Dies kann es ermöglichen, die Anzahl der Stichproben pro Batch zu reduzieren.
Klicke für weitere Informationen zur Konfiguration der SQD-Optionen:
Eine alternative Strategie besteht darin, mehr Stichproben pro Batch zu verwenden, was sicherstellt, dass die meisten der anfänglichen LUCJ-Stichproben im richtigen Teilchenraum während des S-CORE-Verfahrens verwendet werden und einzelne Unterräume eine ausreichende Vielfalt an Elektronenkonfigurationen umfassen. Dies reduziert wiederum die Anzahl der erforderlichen S-CORE-Schritte, wobei nur zwei oder drei Iterationen von SQD benötigt werden, wenn die Anzahl der Stichproben pro Batch groß genug ist. Mehr Stichproben pro Batch führen jedoch zu höheren Rechenkosten für jeden Diagonalisierungsschritt. Daher kannst du das Gleichgewicht zwischen Genauigkeit und Rechenkosten in SQD-Simulationen durch optimale Wahl von sqd_iterations, number_of_batches und samples_per_batch erreichen.
Die SQD IEF-PCM-Studie zeigt, dass bei Verwendung von drei Iterationen von S-CORE die chemische Genauigkeit erreicht werden kann, indem man diese Richtlinien befolgt:
- 600 Stichproben pro Batch in Methanol-SQD IEF-PCM (14e,12o) Simulationen
- 1500 Stichproben pro Batch in Methylamin-SQD IEF-PCM (14e,13o) Simulationen
- 6000 Stichproben pro Batch in Wasser-SQD IEF-PCM (8e,23o) Simulationen
- 16000 Stichproben pro Batch in Ethanol-SQD IEF-PCM (20e,18o) Simulationen
Genau wie die erforderliche Anzahl von Shots in LUCJ ist die erforderliche Anzahl von Stichproben pro Batch, die im S-CORE-Verfahren verwendet wird, stark system- und hardwareabhängig. Die obigen Beispiele können verwendet werden, um den Ausgangspunkt für das Benchmarking der erforderlichen Anzahl von Stichproben pro Batch zu schätzen. Das Tutorial zum systematischen Benchmarking der erforderlichen Anzahl von Stichproben pro Batch findest du hier.
Bereitstellen und Ausführen der Vorlagen-SQD IEF-PCM-Funktion
# Added by doQumentation — required packages for this notebook
!pip install -q ffsim numpy pyscf qiskit qiskit-addon-sqd qiskit-ibm-catalog qiskit-ibm-runtime qiskit-serverless solve-solvent
Authentifizierung
Verwende qiskit-ibm-catalog, um dich mit deinem API-Schlüssel (Token) bei QiskitServerless zu authentifizieren, den du im Dashboard der IBM Quantum Platform findest. Dies ermöglicht die Instanziierung des Serverless-Clients zum Hochladen oder Ausführen der ausgewählten Funktion:
from qiskit_ibm_catalog import QiskitServerless
serverless = QiskitServerless(
channel="ibm_quantum_platform",
instance="INSTANCE_CRN",
token="YOUR_API_KEY" # Use the 44-character API_KEY you created and saved from the IBM Quantum Platform Home dashboard
)
Optional kannst du save_account() verwenden, um deine Anmeldedaten in einer lokalen Umgebung zu speichern (siehe den Leitfaden Richte dein IBM Cloud-Konto ein). Beachte, dass dies deine Anmeldedaten in dieselbe Datei wie QiskitRuntimeService.save_account() schreibt:
QiskitServerless.save_account(token="YOUR_API_KEY", channel="ibm_quantum_platform", instance="INSTANCE_CRN")
Wenn das Konto gespeichert ist, ist es nicht erforderlich, den Token zur Authentifizierung bereitzustellen:
from qiskit_ibm_catalog import QiskitServerless
serverless = QiskitServerless()
Vorlage hochladen
Um eine benutzerdefinierte Qiskit Function hochzuladen, musst du zunächst ein QiskitFunction-Objekt instanziieren, das den Funktionsquellcode definiert. Der Titel ermöglicht es dir, die Funktion zu identifizieren, sobald sie sich im Remote-Cluster befindet. Der Haupteinstiegspunkt ist die Datei, die if __name__ == "__main__" enthält. Wenn dein Workflow zusätzliche Quelldateien benötigt, kannst du ein Arbeitsverzeichnis definieren, das zusammen mit dem Einstiegspunkt hochgeladen wird.
from qiskit_ibm_catalog import QiskitFunction
template = QiskitFunction(
title="sqd_pcm_template",
entrypoint="sqd_pcm_entrypoint.py",
working_dir="./source_files/", # all files in this directory will be uploaded
dependencies=[
"ffsim==0.0.54",
"pyscf==2.9.0",
"qiskit_addon_sqd==0.10.0",
],
)
print(template)
QiskitFunction(sqd_pcm_template)
Sobald die Instanz bereit ist, lade sie auf Serverless hoch:
serverless.upload(template)
QiskitFunction(sqd_pcm_template)
Um zu überprüfen, ob das Programm erfolgreich hochgeladen wurde, nutze serverless.list():
serverless.list()
[QiskitFunction(sqd_pcm_template),
QiskitFunction(hamiltonian_simulation_template)]
Vorlage remote laden und ausführen
Die Funktionsvorlage wurde hochgeladen, sodass du sie remote mit Qiskit Serverless ausführen kannst. Lade zunächst die Vorlage nach Namen:
template = serverless.load("sqd_pcm_template")
print(template)
QiskitFunction(sqd_pcm_template)
Führe als Nächstes die Vorlage mit den Eingaben auf Domänenebene für SQD-IEF PCM aus. Dieses Beispiel spezifiziert einen auf Methanol basierenden Workload.
molecule = {
"atom": """
O -0.04559 -0.75076 -0.00000;
C -0.04844 0.65398 -0.00000;
H 0.85330 -1.05128 -0.00000;
H -1.08779 0.98076 -0.00000;
H 0.44171 1.06337 0.88811;
H 0.44171 1.06337 -0.88811
""", # Must be specified
"basis": "cc-pvdz", # default is "sto-3g"
"spin": 0, # default is 0
"charge": 0, # default is 0
"verbosity": 0, # default is 0
"number_of_active_orb": 12, # Must be specified
"number_of_active_alpha_elec": 7, # Must be specified
"number_of_active_beta_elec": 7, # Must be specified
"avas_selection": [
"%d O %s" % (k, x) for k in [0] for x in ["2s", "2px", "2py", "2pz"]
]
+ ["%d C %s" % (k, x) for k in [1] for x in ["2s", "2px", "2py", "2pz"]]
+ ["%d H 1s" % k for k in [2, 3, 4, 5]], # default is None
}
solvent_options = {
"method": "IEF-PCM", # other available methods are COSMO, C-PCM, SS(V)PE, see https://manual.q-chem.com/5.4/topic_pcm-em.html
"eps": 78.3553, # value for water
}
lucj_options = {
"initial_layout": [
0,
14,
18,
19,
20,
33,
39,
40,
41,
53,
60,
61,
2,
3,
4,
15,
22,
23,
24,
34,
43,
44,
45,
54,
],
"dynamical_decoupling_choice": True,
"twirling_choice": True,
"number_of_shots": 200000,
"optimization_level": 2,
}
sqd_options = {
"sqd_iterations": 3,
"number_of_batches": 10,
"samples_per_batch": 1000,
"max_davidson_cycles": 200,
}
backend_name = "ibm_sherbrooke"
job = template.run(
backend_name=backend_name,
molecule=molecule,
solvent_options=solvent_options,
lucj_options=lucj_options,
sqd_options=sqd_options,
)
print(job.job_id)
39f8fb70-79b2-43ca-b723-84e6b6135821
Überprüfe den detaillierten Status des Jobs:
import time
t0 = time.time()
status = job.status()
if status == "QUEUED":
print(f"time = {time.time()-t0:.2f}, status = QUEUED")
while True:
status = job.status()
if status == "QUEUED":
continue
print(f"time = {time.time()-t0:.2f}, status = {status}")
if status == "DONE" or status == "ERROR":
break
time = 2.35, status = DONE
Während der Job läuft, kannst du Protokolle abrufen, die aus den logger.info-Ausgaben erstellt wurden. Diese können umsetzbare Informationen über den Fortschritt des SQD IEF-PCM-Workflows liefern. Zum Beispiel die gleichen Spinorbitalverbindungen oder die Zwei-Qubit-Tiefe des finalen ISA-Schaltkreises, der zur Ausführung auf Hardware vorgesehen ist.
print(job.logs())
Der Aufruf des Job-Ergebnisses blockiert den Rest des Programms, bis ein Ergebnis verfügbar ist. Nachdem der Job abgeschlossen ist, kannst du die Ergebnisse abrufen. Diese umfassen die freie Solvatationsenergie sowie Informationen über den Batch mit der niedrigsten Energie, den niedrigsten Energiewert und andere nützliche Informationen wie die Gesamtdauer des Solvers.
result = job.result()
result
{'total_energy_hist': array([[-115.14768518, -115.1368396 , -114.19181692, -115.13745429,
-115.1445012 , -114.19673326, -115.1547003 , -114.20563866,
-115.13748344, -115.14764974],
[-115.15768392, -115.15850126, -115.15857275, -115.15770916,
-115.15801684, -115.15822125, -115.15833521, -115.15844051,
-115.15735538, -115.15862354],
[-115.15795148, -115.15847925, -115.15856677, -115.15811156,
-115.15815602, -115.15785171, -115.1583672 , -115.1585533 ,
-115.15833528, -115.15808791]]),
'spin_squared_value_hist': array([[5.37327508e-03, 1.32981759e-02, 1.36214922e-02, 8.84413615e-03,
7.26723578e-03, 1.94875195e-02, 3.03153152e-03, 6.07543106e-03,
1.04951849e-02, 5.36529204e-03],
[6.39397528e-04, 1.36814350e-04, 9.09054260e-05, 5.99361358e-04,
3.64261739e-04, 2.54905866e-04, 2.32540370e-04, 1.53181990e-04,
7.23519739e-04, 6.80737671e-05],
[4.53776416e-04, 1.63043449e-04, 1.05317263e-04, 3.82912836e-04,
3.41047803e-04, 5.18620393e-04, 2.06819142e-04, 1.17086537e-04,
2.32357159e-04, 4.26071537e-04]]),
'solvation_free_energy_hist': array([[-0.00725018, -0.00743955, -0.01132905, -0.0073377 , -0.00722221,
-0.01136705, -0.00719279, -0.01072829, -0.00733404, -0.00725961],
[-0.00719252, -0.00718315, -0.00718074, -0.00719325, -0.00717703,
-0.00718391, -0.00718354, -0.00717928, -0.00719887, -0.0071801 ],
[-0.00719351, -0.00718255, -0.00718198, -0.00718429, -0.00718349,
-0.00718329, -0.0071882 , -0.00718363, -0.00718549, -0.00718814]]),
'occupancy_hist': [[array([0.99712298, 0.99278936, 0.99083163, 0.97328469, 0.98959809,
0.98922134, 0.720333 , 0.25683194, 0.01939338, 0.02840332,
0.00946988, 0.0327204 ]),
array([0.99712298, 0.99278936, 0.99083163, 0.97328469, 0.98959809,
0.98922134, 0.720333 , 0.25683194, 0.01939338, 0.02840332,
0.00946988, 0.0327204 ])],
[array([0.9959042 , 0.9922607 , 0.99018862, 0.99265843, 0.98927447,
0.9900833 , 0.99403876, 0.00989025, 0.01120814, 0.01137717,
0.01152871, 0.01158725]),
array([0.9959042 , 0.9922607 , 0.99018862, 0.99265843, 0.98927447,
0.9900833 , 0.99403876, 0.00989025, 0.01120814, 0.01137717,
0.01152871, 0.01158725])],
[array([0.99590079, 0.99222193, 0.99016753, 0.99265045, 0.98927264,
0.99007179, 0.99407207, 0.00986684, 0.01125181, 0.01141439,
0.01150733, 0.01160243]),
array([0.99590079, 0.99222193, 0.99016753, 0.99265045, 0.98927264,
0.99007179, 0.99407207, 0.00986684, 0.01125181, 0.01141439,
0.01150733, 0.01160243])]],
'lowest_energy_batch': 2,
'lowest_energy_value': -115.1585667736213,
'solvation_free_energy': -0.007181981952470838,
'sci_solver_total_duration': 493.997501373291,
'metadata': {'resources_usage': {'RUNNING: MAPPING': {'CPU_TIME': 6.080063343048096},
'RUNNING: OPTIMIZING_FOR_HARDWARE': {'CPU_TIME': 1.999896764755249},
'RUNNING: WAITING_FOR_QPU': {'CPU_TIME': 6.2850868701934814},
'RUNNING: EXECUTING_QPU': {'QPU_TIME': 21.639373540878296},
'RUNNING: POST_PROCESSING': {'CPU_TIME': 495.40831995010376}},
'num_iterations_executed': 3}}
Beachte, dass die Ergebnis-Metadaten eine Zusammenfassung der Ressourcennutzung enthalten, die es dir ermöglicht, die für jeden Workload erforderliche QPU- und CPU-Zeit besser abzuschätzen (dieses Beispiel wurde auf einem Dummy-Gerät ausgeführt, sodass die tatsächlichen Ressourcennutzungszeiten abweichen können). Nach Abschluss des Jobs ist die gesamte Protokollausgabe verfügbar.
print(job.logs())
2025-06-27 08:42:41,358 INFO job_manager.py:531 -- Runtime env is setting up.
sqd_pcm_entrypoint.run_function:INFO:2025-06-27 08:42:45,015: Starting runtime service
sqd_pcm_entrypoint.run_function:INFO:2025-06-27 08:42:45,621: Backend: ibm_sherbrooke
sqd_pcm_entrypoint.run_function:INFO:2025-06-27 08:42:46,809: Initializing molecule object
sqd_pcm_entrypoint.run_function:INFO:2025-06-27 08:42:51,599: Performing CCSD
Parsing /tmp/ray/session_2025-06-27_08-42-13_898146_1/runtime_resources/working_dir_files/_ray_pkg_4bc93dcc58c04b91/output_sqd_pcm/2025-06-27_08-42-45.fcidump.txt
Overwritten attributes get_ovlp get_hcore of <class 'pyscf.scf.hf_symm.SymAdaptedRHF'>
/usr/local/lib/python3.11/site-packages/pyscf/gto/mole.py:1293: UserWarning: Function mol.dumps drops attribute energy_nuc because it is not JSON-serializable
warnings.warn(msg)
/usr/local/lib/python3.11/site-packages/pyscf/gto/mole.py:1293: UserWarning: Function mol.dumps drops attribute intor_symmetric because it is not JSON-serializable
warnings.warn(msg)
converged SCF energy = -115.049680672847
E(CCSD) = -115.1519910037652 E_corr = -0.1023103309180226
sqd_pcm_entrypoint.run_function:INFO:2025-06-27 08:42:51,694: Same spin orbital connections: [(0, 1), (1, 2), (2, 3), (3, 4), (4, 5), (5, 6), (6, 7), (7, 8), (8, 9), (9, 10), (10, 11)]
sqd_pcm_entrypoint.run_function:INFO:2025-06-27 08:42:51,694: Opposite spin orbital connections: [(0, 0), (4, 4), (8, 8)]
sqd_pcm_entrypoint.run_function:INFO:2025-06-27 08:42:53,718: Optimization level: 2, ops: OrderedDict([('rz', 2438), ('sx', 1496), ('ecr', 766), ('x', 185), ('measure', 24), ('barrier', 1)]), depth: 391
sqd_pcm_entrypoint.run_function:INFO:2025-06-27 08:42:53,736: Two-qubit gate depth: 94
sqd_pcm_entrypoint.run_function:INFO:2025-06-27 08:42:53,737: Submitting sampler job
sqd_pcm_entrypoint.run_function:INFO:2025-06-27 08:42:54,273: Job ID: d1f5j3lqbivc73ebqpj0
sqd_pcm_entrypoint.run_function:INFO:2025-06-27 08:42:54,313: Job Status: QUEUED
sqd_pcm_entrypoint.run_function:INFO:2025-06-27 08:43:24,813: Starting configuration recovery iteration 0
sqd_pcm_entrypoint.run_function:INFO:2025-06-27 08:43:24,841: Batch 0 subspace dimension: 531441
2025-06-27 08:43:24,844 INFO worker.py:1588 -- Using address 172.17.16.124:6379 set in the environment variable RAY_ADDRESS
2025-06-27 08:43:24,847 INFO worker.py:1723 -- Connecting to existing Ray cluster at address: 172.17.16.124:6379...
2025-06-27 08:43:24,876 INFO worker.py:1908 -- Connected to Ray cluster. View the dashboard at [1m[32mhttp://172.17.16.124:8265 [39m[22m
sqd_pcm_entrypoint.run_function:INFO:2025-06-27 08:43:24,945: Batch 1 subspace dimension: 519841
sqd_pcm_entrypoint.run_function:INFO:2025-06-27 08:43:24,950: Batch 2 subspace dimension: 543169
sqd_pcm_entrypoint.run_function:INFO:2025-06-27 08:43:24,955: Batch 3 subspace dimension: 532900
sqd_pcm_entrypoint.run_function:INFO:2025-06-27 08:43:24,960: Batch 4 subspace dimension: 534361
sqd_pcm_entrypoint.run_function:INFO:2025-06-27 08:43:24,964: Batch 5 subspace dimension: 531441
sqd_pcm_entrypoint.run_function:INFO:2025-06-27 08:43:24,969: Batch 6 subspace dimension: 540225
sqd_pcm_entrypoint.run_function:INFO:2025-06-27 08:43:24,974: Batch 7 subspace dimension: 524176
sqd_pcm_entrypoint.run_function:INFO:2025-06-27 08:43:24,979: Batch 8 subspace dimension: 537289
sqd_pcm_entrypoint.run_function:INFO:2025-06-27 08:43:24,983: Batch 9 subspace dimension: 540225
sqd_pcm_entrypoint.run_function:INFO:2025-06-27 08:48:09,006: Lowest energy batch: 6
sqd_pcm_entrypoint.run_function:INFO:2025-06-27 08:48:09,007: Lowest energy value: -115.15470029849135
sqd_pcm_entrypoint.run_function:INFO:2025-06-27 08:48:09,007: Corresponding g_solv value: -0.0071927910374866375
sqd_pcm_entrypoint.run_function:INFO:2025-06-27 08:48:09,007: -----------------------------------
sqd_pcm_entrypoint.run_function:INFO:2025-06-27 08:48:09,007: Starting configuration recovery iteration 1
sqd_pcm_entrypoint.run_function:INFO:2025-06-27 08:48:40,564: Batch 0 subspace dimension: 413449
sqd_pcm_entrypoint.run_function:INFO:2025-06-27 08:48:40,572: Batch 1 subspace dimension: 399424
sqd_pcm_entrypoint.run_function:INFO:2025-06-27 08:48:40,578: Batch 2 subspace dimension: 438244
sqd_pcm_entrypoint.run_function:INFO:2025-06-27 08:48:40,583: Batch 3 subspace dimension: 422500
sqd_pcm_entrypoint.run_function:INFO:2025-06-27 08:48:40,589: Batch 4 subspace dimension: 409600
sqd_pcm_entrypoint.run_function:INFO:2025-06-27 08:48:40,596: Batch 5 subspace dimension: 404496
sqd_pcm_entrypoint.run_function:INFO:2025-06-27 08:48:40,601: Batch 6 subspace dimension: 410881
sqd_pcm_entrypoint.run_function:INFO:2025-06-27 08:48:40,605: Batch 7 subspace dimension: 442225
sqd_pcm_entrypoint.run_function:INFO:2025-06-27 08:48:40,611: Batch 8 subspace dimension: 409600
sqd_pcm_entrypoint.run_function:INFO:2025-06-27 08:48:40,618: Batch 9 subspace dimension: 405769
sqd_pcm_entrypoint.run_function:INFO:2025-06-27 08:49:54,917: Lowest energy batch: 9
sqd_pcm_entrypoint.run_function:INFO:2025-06-27 08:49:54,917: Lowest energy value: -115.15862353596414
sqd_pcm_entrypoint.run_function:INFO:2025-06-27 08:49:54,917: Corresponding g_solv value: -0.0071800982859467006
sqd_pcm_entrypoint.run_function:INFO:2025-06-27 08:49:54,918: -----------------------------------
sqd_pcm_entrypoint.run_function:INFO:2025-06-27 08:49:54,918: Starting configuration recovery iteration 2
sqd_pcm_entrypoint.run_function:INFO:2025-06-27 08:50:25,501: Batch 0 subspace dimension: 399424
sqd_pcm_entrypoint.run_function:INFO:2025-06-27 08:50:25,508: Batch 1 subspace dimension: 412164
sqd_pcm_entrypoint.run_function:INFO:2025-06-27 08:50:25,514: Batch 2 subspace dimension: 432964
sqd_pcm_entrypoint.run_function:INFO:2025-06-27 08:50:25,519: Batch 3 subspace dimension: 400689
sqd_pcm_entrypoint.run_function:INFO:2025-06-27 08:50:25,524: Batch 4 subspace dimension: 432964
sqd_pcm_entrypoint.run_function:INFO:2025-06-27 08:50:25,529: Batch 5 subspace dimension: 418609
sqd_pcm_entrypoint.run_function:INFO:2025-06-27 08:50:25,533: Batch 6 subspace dimension: 418609
sqd_pcm_entrypoint.run_function:INFO:2025-06-27 08:50:25,538: Batch 7 subspace dimension: 425104
sqd_pcm_entrypoint.run_function:INFO:2025-06-27 08:50:25,543: Batch 8 subspace dimension: 404496
sqd_pcm_entrypoint.run_function:INFO:2025-06-27 08:50:25,548: Batch 9 subspace dimension: 429025
sqd_pcm_entrypoint.run_function:INFO:2025-06-27 08:51:37,900: Lowest energy batch: 2
sqd_pcm_entrypoint.run_function:INFO:2025-06-27 08:51:37,900: Lowest energy value: -115.1585667736213
sqd_pcm_entrypoint.run_function:INFO:2025-06-27 08:51:37,901: Corresponding g_solv value: -0.007181981952470838
sqd_pcm_entrypoint.run_function:INFO:2025-06-27 08:51:37,901: -----------------------------------
sqd_pcm_entrypoint.run_function:INFO:2025-06-27 08:51:37,901: SCI_solver totally takes: 493.997501373291 seconds
Nächste Schritte
- Lies den Leitfaden zum Erstellen einer Funktionsvorlage für Hamilton-Simulation
- Schau dir die Quelldateien für diese Vorlage auf GitHub an
Literatur
[1] Danil Kaliakin, Akhil Shajan, Fangchun Liang, and Kenneth M. Merz Jr. Implicit Solvent Sample-Based Quantum Diagonalization, The Journal of Physical Chemistry B, 2025, DOI: 10.1021/acs.jpcb.5c01030
[2] Javier Robledo-Moreno, et al., Chemistry Beyond Exact Solutions on a Quantum-Centric Supercomputer, arXiv:2405.05068 [quant-ph].
[3] Jeffery Yu, et al., Quantum-Centric Algorithm for Sample-Based Krylov Diagonalization, arXiv:2501.09702 [quant-ph].
[4] Keita Kanno, et al., Quantum-Selected Configuration Interaction: classical diagonalization of Hamiltonians in subspaces selected by quantum computers, arXiv:2302.11320 [quant-ph].
[5] Kenji Sugisaki, et al., Hamiltonian simulation-based quantum-selected configuration interaction for large-scale electronic structure calculations with a quantum computer, arXiv:2412.07218 [quant-ph].
[6] Mathias Mikkelsen, Yuya O. Nakagawa, Quantum-selected configuration interaction with time-evolved state, arXiv:2412.13839 [quant-ph].
[7] Herbert, John M. Dielectric continuum methods for quantum chemistry. WIREs Computational Molecular Science, 2021, 11, 1759-0876.
[8] Saki, A. A.; Barison, S.; Fuller, B.; Garrison, J. R.; Glick, J. R.; Johnson, C.; Mezzacapo, A.; Robledo-Moreno, J.; Rossmannek, M.; Schweigert, P. et al. Qiskit addon: sample-based quantum diagonalization, 2024; https://github.com/Qiskit/qiskit-addon-sqd
[9] Asun, Q.; Zhang, X.; Banerjee, S.; Bao, P.; Barbry, M.; Blunt, N. S.; Bogdanov, N. A.; Booth, G. H.; Chen, J.; Cui, Z.-H. PySCF: Python-based Simulations of Chemistry Framework, 2025; https://github.com/pyscf/pyscf
[10] Kevin J. Sung; et al., FFSIM: Faster simulations of fermionic quantum circuits, 2024. https://github.com/qiskit-community/ffsim
%%writefile ./source_files/__init__.py
%%writefile ./source_files/solve_solvent.py
# This code is part of a Qiskit project.
#
# (C) Copyright IBM and Cleveland Clinic 2025
#
# This code is licensed under the Apache License, Version 2.0. You may
# obtain a copy of this license in the LICENSE.txt file in the root directory
# of this source tree or at http://www.apache.org/licenses/LICENSE-2.0.
#
# Any modifications or derivative works of this code must retain this
# copyright notice, and modified files need to carry a notice indicating
# that they have been altered from the originals.
"""Functions for the study of fermionic systems."""
from __future__ import annotations
import warnings
import numpy as np
# DSK Add imports needed for CASCI wrapper
from pyscf import ao2mo, scf, fci
from pyscf.mcscf import avas, casci
from pyscf.solvent import pcm
from pyscf.lib import chkfile, logger
from qiskit_addon_sqd.fermion import (
SCIState,
bitstring_matrix_to_ci_strs,
_check_ci_strs,
)
# DSK Below is the modified CASCI kernel compatible with SQD.
# It utilizes the "fci.selected_ci.kernel_fixed_space"
# as well as enables passing the "batch" and "max_davidson"
# input arguments from "solve_solvent".
# The "batch" contains the CI addresses corresponding to subspaces
# derived from LUCJ and S-CORE calculations.
# The "max_davidson" controls the maximum number of cycles of Davidson's algorithm.
# pylint: disable = unused-argument
def kernel(casci_object, mo_coeff=None, ci0=None, verbose=logger.NOTE, envs=None):
"""CASCI solver compatible with SQD.
Args:
casci_object: CASCI or CASSCF object.
In case of SQD, only CASCI instance is currently incorporated.
mo_coeff : ndarray
orbitals to construct active space Hamiltonian.
In context of SQD, these are either AVAS mo_coeff
or all of the MOs (with option to exclude core MOs).
ci0 : ndarray or custom types FCI solver initial guess.
For SQD the usage of ci0 was not tested.
For external FCI-like solvers, it can be
overloaded different data type. For example, in the state-average
FCI solver, ci0 is a list of ndarray. In other solvers such as
DMRGCI solver, SHCI solver, ci0 are custom types.
kwargs:
envs: dict
In case of SQD this option was not explored,
but in principle this can facilitate the incorporation of the external solvers.
The variable envs is created (for PR 807) to passes MCSCF runtime
environment variables to SHCI solver. For solvers which do not
need this parameter, a kwargs should be created in kernel method
and "envs" pop in kernel function.
"""
if mo_coeff is None:
mo_coeff = casci_object.mo_coeff
if ci0 is None:
ci0 = casci_object.ci
log = logger.new_logger(casci_object, verbose)
t0 = (logger.process_clock(), logger.perf_counter())
log.debug("Start CASCI")
ncas = casci_object.ncas
nelecas = casci_object.nelecas
# The start of SQD version of kernel
# DSK add the read of configurations for batch
ci_strs_sqd = casci_object.batch
# DSK add the input for the maximum number of cycles of Davidson's algorithm
max_davidson = casci_object.max_davidson
# DSK add electron up and down count and norb = ncas
n_up = nelecas[0]
n_dn = nelecas[1]
norb = ncas
# DSK Eigenstate solver info
sqd_verbose = verbose
# DSK ERI read
eri_cas = ao2mo.restore(1, casci_object.get_h2eff(), casci_object.ncas)
t1 = log.timer("integral transformation to CAS space", *t0)
# DSK 1e integrals
h1eff, energy_core = casci_object.get_h1eff()
log.debug("core energy = %.15g", energy_core)
t1 = log.timer("effective h1e in CAS space", *t1)
if h1eff.shape[0] != ncas:
raise RuntimeError(
"Active space size error. nmo=%d ncore=%d ncas=%d" # pylint: disable=consider-using-f-string
% (mo_coeff.shape[1], casci_object.ncore, ncas)
)
# DSK fcisolver needs to be defined in accordance with SQD
# in this software stack it is done in the "solve_solvent" portion of the code.
myci = casci_object.fcisolver
e_cas, sqdvec = fci.selected_ci.kernel_fixed_space(
myci,
h1eff,
eri_cas,
norb,
(n_up, n_dn),
ci_strs=ci_strs_sqd,
verbose=sqd_verbose,
max_cycle=max_davidson,
)
# DSK fcivec is the general name for CI vector assigned by PySCF.
# Depending on type of solver it is either FCI or SCI vector.
# In case of sqd we can call it "sqdvec" for clarity.
# Nonetheless, for further processing PySCF expects
# this data structure to be called fcivec, regardless of the used solver.
fcivec = sqdvec
t1 = log.timer("CI solver", *t1)
e_tot = energy_core + e_cas
# Returns either standard CASCI data or SQD data. Return depends on "sqd_run" True/False.
return e_tot, e_cas, fcivec
# Replace standard CASCI kernel with the SQD-compatible CASCI kernel defined above
casci.kernel = kernel
def solve_solvent(
bitstring_matrix: tuple[np.ndarray, np.ndarray] | np.ndarray,
/,
myeps: float,
mysolvmethod: str,
myavas: list,
num_orbitals: int,
*,
spin_sq: int | None = None,
max_davidson: int = 100,
verbose: int | None = 0,
checkpoint_file: str,
) -> tuple[float, SCIState, list[np.ndarray], float]:
"""Approximate the ground state given molecular integrals and a set of electronic configurations.
Args:
bitstring_matrix: A set of configurations defining the subspace onto which the Hamiltonian
will be projected and diagonalized. This is a 2D array of ``bool`` representations of bit
values such that each row represents a single bitstring. The spin-up configurations
should be specified by column indices in range ``(N, N/2]``, and the spin-down
configurations should be specified by column indices in range ``(N/2, 0]``, where ``N``
is the number of qubits.
(DEPRECATED) The configurations may also be specified by a length-2 tuple of sorted 1D
arrays containing unsigned integer representations of the determinants. The two lists
should represent the spin-up and spin-down orbitals, respectively.
To build PCM model PySCF needs the structure of the molecule. Hence, the electron integrals
(hcore and eri) are not enough to form IEF-PCM simulation. Instead the "start.chk" file is used.
This workflow also requires additional information about solute and solvent,
which is reflected by additional arguments below
myeps: Dielectric parameter of the solvent.
mysolvmethod: Solvent model, which can be IEF-PCM, COSMO, C-PCM, SS(V)PE,
see https://manual.q-chem.com/5.4/topic_pcm-em.html
At the moment only IEF-PCM was tested.
In principle two other models from PySCF "solvent" module can be used as well,
namely SMD and polarizable embedding (PE).
The SMD and PE were not tested yet and their usage requires addition of more
input arguments for "solve_solvent".
myavas: This argument allows user to select active space in solute with AVAS.
The corresponding list should include target atomic orbitals.
If myavas=None, then active space selected based on number of orbitals
derived from ci_strs.
It is assumed that if myavas=None, then the target calculation is either
a) corresponds to full basis case.
b) close to full basis case and only few core orbitals are excluded.
num_orbitals: Number of orbitals, which is essential when myavas = None.
In AVAS case number of orbitals and electrons is derived by AVAS procedure itself.
spin_sq: Target value for the total spin squared for the ground state.
If ``None``, no spin will be imposed.
max_davidson: The maximum number of cycles of Davidson's algorithm
verbose: A verbosity level between 0 and 10
checkpoint_file: Name of the checkpoint file
NOTE: For now open shell functionality is not supported in SQD PCM calculations.
Hence, at the moment solve_solvent does not include open_shell as one of the arguments.
Returns:
- Minimum energy from SCI calculation
- The SCI ground state
- Average occupancy of the alpha and beta orbitals, respectively
- Expectation value of spin-squared
- Solvation free energy
"""
# Unlike the "solve_fermion", the "solve_solvent" utilizes the "checkpoint" file to
# get the starting HF information, which means that "solve_solvent" does not accept
# "hcore" and "eri" as the input arguments.
# Instead "hcore" and "eri" are generated inside of the custom SQD-compatible
# CASCI kernel (defined above).
# The generation of "hcore" and "eri" is based on the information from "checkpoint" file
# as well as "myavas" and "num_orbitals" input arguments.
# DSK this part handles addresses and is identical to "solve_fermion"
if isinstance(bitstring_matrix, tuple):
warnings.warn(
"Passing the input determinants as integers is deprecated. "
"Users should instead pass a bitstring matrix defining the subspace.",
DeprecationWarning,
stacklevel=2,
)
ci_strs = bitstring_matrix
else:
# This will become the default code path after the deprecation period.
ci_strs = bitstring_matrix_to_ci_strs(bitstring_matrix, open_shell=False)
ci_strs = _check_ci_strs(ci_strs)
num_up = format(ci_strs[0][0], "b").count("1")
num_dn = format(ci_strs[1][0], "b").count("1")
# DSK assign verbosity
verbose_ci = verbose
# DSK add information about solute and solvent.
# Since PCM model needs the information about the structure of the molecule
# one cannot use only FCIDUMP. Instead converged HF data can be passed from "checkpoint" file
# along with "mol" object containing the geometry and other information about the solute.
############################################
# This section is specific to "solve_solvent" and is not present in "solve_fermion".
# In case of "solve_fermion" the "eri" and "hcore" are passed directly to
# "fci.selected_ci.kernel_fixed_space".
# In case of "solve_solvent" the incorporation of the polarizable continuum model
# requires utilization of "CASCI.with_solvent"
# data object from PySCF, where underlying CASCI.base_kernel has to be replaced
# with SQD-compatible version.
# Due to these differences in the implementation the "solve_solvent" recovers
# the converged mean field results and "molecule" object from "checkpoint" file
# (instead of using FCIDUMP),
# followed by passing of solute, solvent, and active space information to "CASCI.with_solvent".
# This includes the initiation of "mol", "cm", "mf", and "mc" data structures.
mol = chkfile.load_mol(checkpoint_file)
# DSK Initiation of the solvent model
cm = pcm.PCM(mol)
cm.eps = myeps # solute eps value
cm.method = mysolvmethod # IEF-PCM, COSMO, C-PCM, SS(V)PE,
# see https://manual.q-chem.com/5.4/topic_pcm-em.html
# DSK Read-in converged RHF solution
scf_result_dic = chkfile.load(checkpoint_file, "scf")
mf = scf.RHF(mol).PCM(cm)
mf.__dict__.update(scf_result_dic)
# Identify the active space based on the user input of AVAS or number of orbitals and electrons
if myavas is not None:
orbs = myavas
avas_obj = avas.AVAS(mf, orbs, with_iao=True)
avas_obj.kernel()
ncas, nelecas, _, _, _ = (
avas_obj.ncas,
avas_obj.nelecas,
avas_obj.mo_coeff,
avas_obj.occ_weights,
avas_obj.vir_weights,
)
else:
ncas = num_orbitals
nelecas = (num_up, num_dn)
# Initiate the "CASCI.with_solvent" object
mc = casci.CASCI(mf, ncas=ncas, nelecas=nelecas).PCM(cm)
# Replace mo_coeff with ones produced by AVAS if AVAS is utilized
if myavas is not None:
mc.mo_coeff = avas_obj.mo_coeff
# Read-in the configuration interaction subspace derived from LUCJ and S-CORE
mc.batch = ci_strs
# Assign number of maximum Davidson steps
mc.max_davidson = max_davidson
####### The definition of "fcisolver" object is identical to "solve_fermion" case ########
myci = fci.selected_ci.SelectedCI()
if spin_sq is not None:
myci = fci.addons.fix_spin_(myci, ss=spin_sq)
mc.fcisolver = myci
mc.verbose = verbose_ci
#########################################################################################
# Initiate the "CASCI.with_solvent" simulation with SQD-compatible based CASCI kernel.
mc_result = mc.kernel()
# Get data out of the "CASCI.with_solvent" object
e_sci = mc_result[0]
sci_vec = mc_result[2]
# Here we get additional output comparing to "solve_fermion",
# which is the solvation free energy (G_solv)
g_solv = mc.with_solvent.e
#####################################################
# The remainder of the code in solve_solvent is nearly identical to solve_fermion code.
# However, there are two exceptions in "solve_solvent":
# 1) The dm2 is currently not computed, but can be included if needed
# 2) e_sci is directly output as the result of CASCI.with_solvent object.
# Hence, the two following lines of code are not present in "solve_solvent"
# comparing to the "solve_fermion" code:
# dm2 = myci.make_rdm2(sci_vec, norb, (num_up, num_dn))
# e_sci = np.einsum("pr,pr->", dm1, hcore) + 0.5 * np.einsum("prqs,prqs->", dm2, eri)
# Calculate the avg occupancy of each orbital
dm1 = myci.make_rdm1s(sci_vec, ncas, (num_up, num_dn))
avg_occupancy = [np.diagonal(dm1[0]), np.diagonal(dm1[1])]
# Compute total spin
spin_squared = myci.spin_square(sci_vec, ncas, (num_up, num_dn))[0]
# Convert the PySCF SCIVector to internal format. We access a private field here,
# so we assert that we expect the SCIVector output from kernel_fixed_space to
# have its _strs field populated with alpha and beta strings.
assert isinstance(sci_vec._strs[0], np.ndarray) and isinstance(sci_vec._strs[1], np.ndarray)
assert sci_vec.shape == (len(sci_vec._strs[0]), len(sci_vec._strs[1]))
sci_state = SCIState(
amplitudes=np.array(sci_vec),
ci_strs_a=sci_vec._strs[0],
ci_strs_b=sci_vec._strs[1],
)
return e_sci, sci_state, avg_occupancy, spin_squared, g_solv
%%writefile ./source_files/sqc_pcm_entrypoint.py
# This code is part of a Qiskit project.
#
# (C) Copyright IBM and Cleveland Clinic 2025
#
# This code is licensed under the Apache License, Version 2.0. You may
# obtain a copy of this license in the LICENSE.txt file in the root directory
# of this source tree or at http://www.apache.org/licenses/LICENSE-2.0.
#
# Any modifications or derivative works of this code must retain this
# copyright notice, and modified files need to carry a notice indicating
# that they have been altered from the originals.
"""
SQD-PCM Function Template source code.
"""
from pathlib import Path
from typing import Any
from datetime import datetime
import os
import sys
import json
import logging
import time
import traceback
import numpy as np
import ffsim
from pyscf import gto, scf, mcscf, ao2mo, tools, cc
from pyscf.lib import chkfile
from pyscf.mcscf import avas
from pyscf.solvent import pcm
from qiskit import QuantumCircuit, QuantumRegister
from qiskit.transpiler.preset_passmanagers import generate_preset_pass_manager
from qiskit.primitives import BackendSamplerV2
from qiskit_addon_sqd.counts import counts_to_arrays
from qiskit_addon_sqd.configuration_recovery import recover_configurations
from qiskit_addon_sqd.fermion import bitstring_matrix_to_ci_strs
from qiskit_addon_sqd.subsampling import postselect_and_subsample
from qiskit_ibm_runtime import QiskitRuntimeService, SamplerV2
from qiskit_serverless import get_arguments, save_result, distribute_task, get, update_status, Job
current_dir = os.path.dirname(os.path.abspath(__file__))
sys.path.insert(0, current_dir)
from solve_solvent import solve_solvent # pylint: disable=wrong-import-position
logger = logging.getLogger(__name__)
def run_function(
backend_name: str,
molecule: dict,
solvent_options: dict,
sqd_options: dict,
lucj_options: dict | None = None,
**kwargs,
) -> dict[str, Any]:
"""
Main entry point for the SQD-PCM (Polarizable Continuum Model) workflow.
This function encapsulates the end-to-end execution of the algorithm.
Args:
backend_name: Identifier for the target backend, required for all
workflows that access IBM Quantum hardware.
molecule: dictionary with molecule information:
- "atom" (str): required field, follows pyscf specification for atomic geometry.
For example, for methanol the value would be::
'''
O -0.04559 -0.75076 -0.00000;
C -0.04844 0.65398 -0.00000;
H 0.85330 -1.05128 -0.00000;
H -1.08779 0.98076 -0.00000;
H 0.44171 1.06337 0.88811;
H 0.44171 1.06337 -0.88811;
'''
- "number_of_active_orb" (int): required field
- "number_of_active_alpha_elec" (int): required field
- "number_of_active_beta_elec" (int): required field
- "basis" (str): optional field, default is "sto-3g"
- "verbosity" (int): optional field, default is 0
- "charge" (int): optional field, default is 0
- "spin" (int): optional field, default is 0
- "avas_selection" (list[str] | None): optional field, default is None
solvent_options: dictionary with solvent options information:
- "method" (str): required field. Method for computing solvent reaction field
for the PCM. Accepted values are: "IEF-PCM", "COSMO",
"C-PCM", "SS(V)PE", see https://manual.q-chem.com/5.4/topic_pcm-em.html
- "eps" (float): required field. Dielectric constant of the solvent in the PCM.
sqd_options: dictionary with sqd options information:
- "sqd_iterations" (int): required field.
- "number_of_batches" (int): required field.
- "samples_per_batch" (int): required field.
- "max_davidson_cycles" (int): required field.
lucj_options: optional dictionary with lucj options information:
- "optimization_level" (int): optional field, default is 2
- "initial_layout" (list[int]): optional field, default is None
- "dynamical_decoupling" (bool): optional field, default is True
- "twirling" (bool): optional field, default is True
- "number_of_shots" (int): optional field, default is 10000
**kwargs
Optional keyword arguments to customize behavior. Existing kwargs include:
- "files_name" (str): optional name for output files (enabled for local testing)
- "testing_backend" (FakeBackendV2): optional fake backend instance to bypass
qiskit runtime service instantiation (enabled for local testing)
- "count_dict_file_name" (str): path to a count dict file to bypass primitive
execution and jump directly to SQD section (enabled for local testing)
Returns:
The function should return the execution results as a dictionary with string keys.
This is to ensure compatibility with ``qiskit_serverless.save_result``.
"""
# Preparation Step: Input validation.
# Do this at the top of the function definition so it fails early if any required
# arguments are missing or invalid.
# Molecule parsing
# Required:
geo = molecule["atom"]
num_active_orb = molecule["number_of_active_orb"]
num_active_alpha = molecule["number_of_active_alpha_elec"]
num_active_beta = molecule["number_of_active_beta_elec"]
# Optional:
input_basis = molecule.get("basis", "sto-3g")
input_verbosity = molecule.get("verbosity", 0)
input_charge = molecule.get("charge", 0)
input_spin = molecule.get("spin", 0)
myavas = molecule.get("avas_selection", None)
# Solvent options parsing
myeps = solvent_options["eps"]
mymethod = solvent_options["method"]
# LUCJ options parsing
if lucj_options is None:
lucj_options = {}
opt_level = lucj_options.get("optimization_level", 2)
initial_layout = lucj_options.get("initial_layout", None)
use_dd = lucj_options.get("dynamical_decoupling", True)
use_twirling = lucj_options.get("twirling", True)
num_shots = lucj_options.get("number_of_shots", True)
# SQD options parsing
iterations = sqd_options["sqd_iterations"]
n_batches = sqd_options["number_of_batches"]
samples_per_batch = sqd_options["samples_per_batch"]
max_davidson_cycles = sqd_options["max_davidson_cycles"]
# kwarg parsing (local testing)
testing_backend = kwargs.get("testing_backend", None)
count_dict_file_name = kwargs.get("count_dict_file_name", None)
files_name = kwargs.get("files_name", datetime.now().strftime("%Y-%m-%d_%H-%M-%S"))
output_path = Path.cwd() / "output_sqd_pcm"
output_path.mkdir(exist_ok=True)
datafiles_name = str(output_path) + "/" + files_name
# --
# Preparation Step: Qiskit Runtime & primitive configuration for
# execution on IBM Quantum hardware.
if testing_backend is None:
# Initialize Qiskit Runtime Service
logger.info("Starting runtime service")
service = QiskitRuntimeService(
channel=os.environ["QISKIT_IBM_CHANNEL"],
instance=os.environ["IBM_CLOUD_INSTANCE"],
token=os.environ["your-API_KEY"], # Use the 44-character API_KEY you created and saved from the IBM Quantum Platform Home dashboard
)
backend = service.backend(backend_name)
logger.info(f"Backend: {backend.name}")
# Set up sampler and corresponding options
sampler = SamplerV2(backend)
sampler.options.dynamical_decoupling.enable = use_dd
sampler.options.twirling.enable_measure = False
sampler.options.twirling.enable_gates = use_twirling
sampler.options.default_shots = num_shots
else:
backend = testing_backend
logger.info(f"Testing backend: {backend.name}")
# Set up backend sampler.
# This doesn't allow running with twirling and dd
sampler = BackendSamplerV2(backend=testing_backend)
# Once the preparation steps are completed, the algorithm can be structured following a
# Qiskit Pattern workflow:
# https://docs.quantum.ibm.com/guides/intro-to-patterns
# --
# Step 1: Map
# In this step, input arguments are used to construct relevant quantum circuits and operators
start_mapping = time.time()
update_status(Job.MAPPING)
# Initialize the molecule object (pyscf)
logger.info("Initializing molecule object")
mol = gto.Mole()
mol.build(
atom=geo,
basis=input_basis,
verbose=input_verbosity,
charge=input_charge,
spin=input_spin,
symmetry=False,
) # Not tested for symmetry calculations
cm = pcm.PCM(mol)
cm.eps = myeps
cm.method = mymethod
mf = scf.RHF(mol).PCM(cm)
# Generation of checkpoint file for the solute and solvent
# which will be used reused in all subsequent sections
checkpoint_file_name = str(datafiles_name + ".chk")
mf.chkfile = checkpoint_file_name
mf.kernel()
# Read-in the information about the molecule
mol = chkfile.load_mol(checkpoint_file_name)
# Read-in RHF data
scf_result_dic = chkfile.load(checkpoint_file_name, "scf")
mf = scf.RHF(mol)
mf.__dict__.update(scf_result_dic)
# LUCJ uses isolated solute
mf.kernel()
# Initialize orbital selection based on user input
if myavas is not None:
orbs = myavas
avas_out = avas.AVAS(mf, orbs, with_iao=True)
avas_out.kernel()
ncas, nelecas = (avas_out.ncas, avas_out.nelecas)
else:
ncas = num_active_orb
nelecas = (
num_active_alpha,
num_active_beta,
)
# LUCJ Step:
# Generate active space
mc = mcscf.CASCI(mf, ncas=ncas, nelecas=nelecas)
if myavas is not None:
mc.mo_coeff = avas_out.mo_coeff
mc.batch = None
# Reliable and most convenient way to do the CCSD on only the active space
# is to create the FCIDUMP file and then run the CCSD calculation only on the
# orbitals stored in the FCIDUMP file.
h1e_cas, ecore = mc.get_h1eff()
h2e_cas = ao2mo.restore(1, mc.get_h2eff(), mc.ncas)
fcidump_file_name = str(datafiles_name + ".fcidump.txt")
tools.fcidump.from_integrals(
fcidump_file_name,
h1e_cas,
h2e_cas,
ncas,
nelecas,
nuc=ecore,
ms=0,
orbsym=[1] * ncas,
)
logger.info("Performing CCSD")
# Read FCIDUMP and perform CCSD on only active space
mf_as = tools.fcidump.to_scf(fcidump_file_name)
mf_as.kernel()
mc_cc = cc.CCSD(mf_as)
mc_cc.kernel()
mc_cc.t1 # pylint: disable=pointless-statement
t2 = mc_cc.t2
n_reps = 2
norb = ncas
if myavas is not None:
nelec = (int(nelecas / 2), int(nelecas / 2))
else:
nelec = nelecas
alpha_alpha_indices = [(p, p + 1) for p in range(norb - 1)]
alpha_beta_indices = [(p, p) for p in range(0, norb, 4)]
logger.info(f"Same spin orbital connections: {alpha_alpha_indices}")
logger.info(f"Opposite spin orbital connections: {alpha_beta_indices}")
# Construct LUCJ op
ucj_op = ffsim.UCJOpSpinBalanced.from_t_amplitudes(
t2, n_reps=n_reps, interaction_pairs=(alpha_alpha_indices, alpha_beta_indices)
)
# Construct circuit
qubits = QuantumRegister(2 * norb, name="q")
circuit = QuantumCircuit(qubits)
circuit.append(ffsim.qiskit.PrepareHartreeFockJW(norb, nelec), qubits)
circuit.append(ffsim.qiskit.UCJOpSpinBalancedJW(ucj_op), qubits)
circuit.measure_all()
end_mapping = time.time()
# --
# Step 2: Optimize
# Transpile circuits to match ISA
start_optimizing = time.time()
update_status(Job.OPTIMIZING_HARDWARE)
pass_manager = generate_preset_pass_manager(
optimization_level=opt_level,
backend=backend,
initial_layout=initial_layout,
)
pass_manager.pre_init = ffsim.qiskit.PRE_INIT
transpiled = pass_manager.run(circuit)
end_optimizing = time.time()
logger.info(
f"Optimization level: {opt_level}, ops: {transpiled.count_ops()}, depth: {transpiled.depth()}"
)
two_q_depth = transpiled.depth(lambda x: x.operation.num_qubits == 2)
logger.info(f"Two-qubit gate depth: {two_q_depth}")
# --
# Step 3: Execute on Hardware
# Submit the underlying Sampler job. Note that this is not the
# actual function job.
if count_dict_file_name is None:
# Submit the LUCJ job
logger.info("Submitting sampler job")
job = sampler.run([transpiled])
logger.info(f"Job ID: {job.job_id()}")
logger.info(f"Job Status: {job.status()}")
start_waiting_qpu = time.time()
while job.status() == "QUEUED":
update_status(Job.WAITING_QPU)
time.sleep(5)
end_waiting_qpu = time.time()
update_status(Job.EXECUTING_QPU)
# Wait until job is complete
result = job.result()
end_executing_qpu = time.time()
pub_result = result[0]
counts_dict = pub_result.data.meas.get_counts()
waiting_qpu_time = end_waiting_qpu - start_waiting_qpu
executing_qpu_time = end_executing_qpu - end_waiting_qpu
else:
# read LUCJ samples from count_dict
logger.info("Skipping sampler, loading counts dict from file")
with open(count_dict_file_name, "r") as file:
count_dict_string = file.read().replace("\n", "")
counts_dict = json.loads(count_dict_string.replace("'", '"'))
waiting_qpu_time = 0
executing_qpu_time = 0
# --
# Step 4: Post-process
start_pp = time.time()
update_status(Job.POST_PROCESSING)
# SQD-PCM section
start = time.time()
# Orbitals, electron, and spin initialization
num_orbitals = ncas
if myavas is not None:
num_elec_a = num_elec_b = int(nelecas / 2)
else:
num_elec_a, num_elec_b = nelecas
spin_sq = input_spin
# Convert counts into bitstring and probability arrays
bitstring_matrix_full, probs_arr_full = counts_to_arrays(counts_dict)
# We set qiskit_serverless to explicitly reserve 1 cpu per thread, as
# the task is CPU-bound and might degrade in performance when sharing
# a core at scale (this might not be the case with smaller examples)
@distribute_task(target={"cpu": 1})
def solve_solvent_parallel(
batches,
myeps,
mysolvmethod,
myavas,
num_orbitals,
spin_sq,
max_davidson,
checkpoint_file,
):
return solve_solvent( # sqd for pyscf
batches,
myeps,
mysolvmethod,
myavas,
num_orbitals,
spin_sq=spin_sq,
max_davidson=max_davidson,
checkpoint_file=checkpoint_file,
)
e_hist = np.zeros((iterations, n_batches)) # energy history
s_hist = np.zeros((iterations, n_batches)) # spin history
g_solv_hist = np.zeros((iterations, n_batches)) # g_solv history
occupancy_hist = []
avg_occupancy = None
num_ran_iter = 0
for i in range(iterations):
logger.info(f"Starting configuration recovery iteration {i}")
# On the first iteration, we have no orbital occupancy information from the
# solver, so we begin with the full set of noisy configurations.
if avg_occupancy is None:
bs_mat_tmp = bitstring_matrix_full
probs_arr_tmp = probs_arr_full
# If we have average orbital occupancy information, we use it to refine the full
# set of noisy configurations
else:
bs_mat_tmp, probs_arr_tmp = recover_configurations(
bitstring_matrix_full, probs_arr_full, avg_occupancy, num_elec_a, num_elec_b
)
# Create batches of subsamples. We post-select here to remove configurations
# with incorrect hamming weight during iteration 0, since no config recovery was performed.
batches = postselect_and_subsample(
bs_mat_tmp,
probs_arr_tmp,
hamming_right=num_elec_a,
hamming_left=num_elec_b,
samples_per_batch=samples_per_batch,
num_batches=n_batches,
)
# Run eigenstate solvers in a loop. This loop should be parallelized for larger problems.
e_tmp = np.zeros(n_batches)
s_tmp = np.zeros(n_batches)
g_solvs_tmp = np.zeros(n_batches)
occs_tmp = []
coeffs = []
res1 = []
for j in range(n_batches):
strs_a, strs_b = bitstring_matrix_to_ci_strs(batches[j])
logger.info(f"Batch {j} subspace dimension: {len(strs_a) * len(strs_b)}")
res1.append(
solve_solvent_parallel(
batches[j],
myeps,
mymethod,
myavas,
num_orbitals,
spin_sq=spin_sq,
max_davidson=max_davidson_cycles,
checkpoint_file=checkpoint_file_name,
)
)
res = get(res1)
for j in range(n_batches):
energy_sci, coeffs_sci, avg_occs, spin, g_solv = res[j]
e_tmp[j] = energy_sci
s_tmp[j] = spin
g_solvs_tmp[j] = g_solv
occs_tmp.append(avg_occs)
coeffs.append(coeffs_sci)
# Combine batch results
avg_occupancy = tuple(np.mean(occs_tmp, axis=0))
# Track optimization history
e_hist[i, :] = e_tmp
s_hist[i, :] = s_tmp
g_solv_hist[i, :] = g_solvs_tmp
occupancy_hist.append(avg_occupancy)
lowest_e_batch_index = np.argmin(e_hist[i, :])
logger.info(f"Lowest energy batch: {lowest_e_batch_index}")
logger.info(f"Lowest energy value: {np.min(e_hist[i, :])}")
logger.info(f"Corresponding g_solv value: {g_solv_hist[i, lowest_e_batch_index]}")
logger.info("-----------------------------------")
num_ran_iter += 1
end_pp = time.time()
end = time.time()
duration = end - start
logger.info(f"SCI_solver totally takes: {duration} seconds")
metadata = {
"resources_usage": {
"RUNNING: MAPPING": {
"CPU_TIME": end_mapping - start_mapping,
},
"RUNNING: OPTIMIZING_FOR_HARDWARE": {
"CPU_TIME": end_optimizing - start_optimizing,
},
"RUNNING: WAITING_FOR_QPU": {
"CPU_TIME": waiting_qpu_time,
},
"RUNNING: EXECUTING_QPU": {
"QPU_TIME": executing_qpu_time,
},
"RUNNING: POST_PROCESSING": {
"CPU_TIME": end_pp - start_pp,
},
},
"num_iterations_executed": num_ran_iter,
}
output = {
"total_energy_hist": e_hist,
"spin_squared_value_hist": s_hist,
"solvation_free_energy_hist": g_solv_hist,
"occupancy_hist": occupancy_hist,
"lowest_energy_batch": lowest_e_batch_index,
"lowest_energy_value": np.min(e_hist[i, :]),
"solvation_free_energy": g_solv_hist[i, lowest_e_batch_index],
"sci_solver_total_duration": duration,
"metadata": metadata,
}
return output
def set_up_logger(my_logger: logging.Logger, level: int = logging.INFO) -> None:
"""Logger setup to communicate logs through serverless."""
log_fmt = "%(module)s.%(funcName)s:%(levelname)s:%(asctime)s: %(message)s"
formatter = logging.Formatter(log_fmt)
# Set propagate to `False` since handlers are to be attached.
my_logger.propagate = False
stream_handler = logging.StreamHandler()
stream_handler.setFormatter(formatter)
my_logger.addHandler(stream_handler)
my_logger.setLevel(level)
# This is the section where `run_function` is called, it's boilerplate code and can be used
# without customization.
if __name__ == "__main__":
# Use serverless helper function to extract input arguments,
input_args = get_arguments()
# Allow to configure logging level
logging_level = input_args.get("logging_level", logging.INFO)
set_up_logger(logger, logging_level)
try:
func_result = run_function(**input_args)
# Use serverless function to save the results that
# will be returned in the job.
save_result(func_result)
except Exception:
save_result(traceback.format_exc())
raise
sys.exit(0)
# This cell is hidden from users. It verifies both source listings are identical then deletes the working folder we created
import shutil
with open("./source_files/sqd_pcm_entrypoint.py") as f1:
with open("./source_files/sqd_pcm_entrypoint.py") as f2:
assert f1.read() == f2.read()
with open("./source_files/solve_solvent.py") as f1:
with open("./source_files/solve_solvent.py") as f2:
assert f1.read() == f2.read()
with open("./source_files/__init__.py") as f1:
with open("./source_files/__init__.py") as f2:
assert f1.read() == f2.read()
shutil.rmtree("./source_files/")