The Suzuki-Miyaura cross-coupling reaction represents one of the cornerstone transformations of modern catalysis for the construction of carbon-carbon bonds. Despite its industrial and scientific centrality, most studies report yield as the sole performance metric, reducing an intrinsically dynamic process to a static end-point measurement. This simplification obscures the rich information embedded in concentration-time trajectories, which encode transient regimes, stiff dynamics, and the impact of ligands and base on catalytically relevant intermediates that are not experimentally observable. This thesis combines mechanistic simulation, hybrid deep learning for dynamical systems, and structural signal extraction from ligands to better understand the role of the catalytic system in governing reaction behavior. In the first part, ODE models derived from kinetic studies of increasing complexity are transformed into generative simulators to explore operational spaces via Latin Hypercube Sampling. The resulting in silico trajectories incorporate controlled noise, realistically reproducing experimental uncertainty. Sequential models (LSTMs) and continuous models constrained by chemical structure (stoichiometrically consistent Neural ODEs) are trained on these data, with loss functions restricted to experimentally measurable species. The results demonstrate that neural surrogates can learn latent catalytic dynamics and reproduce coherent behaviors even without supervision on palladium intermediates, outlining a new framework for hybrid modeling in catalysis. Subsequently, a retrospective extraction of structure-activity signals is performed using heterogeneous literature data, following rigorous curation of a reaction database containing catalyst-ligand complexes, in which ligands are converted into canonical SMILES. An Explainable Boosting Machine models yield as a function of operating conditions, isolating a residual signal corrected for these conditions. This signal is then subjected to motif mining using Morgan fingerprints and non-parametric statistical testing. Overall, the thesis proposes a reproducible computational platform that: (i) redefines kinetic modeling of the Suzuki-Miyaura reaction through structurally consistent and physically informed neural surrogates; and (ii) provides guidelines for the design of prospective datasets and catalysis studies oriented toward causal inference.
La reazione di cross-coupling di Suzuki-Miyaura costituisce una delle trasformazioni cardine della catalisi moderna per la costruzione di legami carbonio-carbonio. Nonostante la sua centralità industriale e scientifica, la maggior parte degli studi riporta la resa come unica metrica di prestazione, riducendo un processo intrinsecamente dinamico a una misura statica di fine reazione. Questa semplificazione eclissa la ricchezza informativa contenuta nelle traiettorie concentrazione-tempo che codificano regimi transitori, fenomeni di stiff dynamics e l’impatto dei leganti e della base su intermedi catalitici non osservabili sperimentalmente. Questa tesi combina quindi simulazione meccanicistica, modelli ibridi di deep learning per sistemi dinamici ed estrazione di segnali strutturali dei leganti al fine di comprendere meglio il ruolo del sistema catalitico sulla reazione. Nella prima parte, modelli ODE derivati da studi cinetici di complessità crescente vengono trasformati in simulatori generativi per esplorare spazi operativi tramite Latin Hypercube Sampling. Le traiettorie in silico risultanti inclusive di rumore controllato riproducono, quindi, in maniera realistica l’incertezza sperimentale. Su tali dati vengono addestrati modelli sequenziali (LSTM) e modelli continui vincolati dalla struttura chimica (NeuralODE stechiometricamente consistenti), con funzioni di loss limitate alle sole specie misurabili. I risultati dimostrano che surrogati neurali possono apprendere dinamiche catalitiche latenti e riprodurre comportamenti coerenti anche in assenza di supervisione sugli intermedi al palladio, delineando un nuovo quadro per la modellazione ibrida in catalisi. Successivamente, viene effettuata un’estrazione retrospettiva di segnali struttura-attività a partire da dati eterogenei di letteratura, previa rigorosa revisione di un database di reazioni contenente complessi catalizzatore-legante, in cui i leganti sono stati convertiti in SMILES canonici. Un Explainable Boosting Machine modella la resa in funzione delle condizioni operative, isolando un segnale residuo corretto per tali condizioni. Questo segnale viene quindi sottoposto a motif mining mediante Morgan fingerprints e test statistici non parametrici. Nel complesso, la tesi propone una piattaforma computazionale riproducibile che: (i) ridefinisce la modellazione cinetica nella reazione Suzuki-Miyaura attraverso surrogati neurali strutturalmente coerenti e fisicamente informati; e (ii) fornisce linee guida per la progettazione di dataset prospettici e studi catalitici orientati all’inferenza causale.
Hybrid neural modeling and ligand signal mining in Suzuki-Miyaura cross-coupling reactions
Catalani, Ivan
2024/2025
Abstract
The Suzuki-Miyaura cross-coupling reaction represents one of the cornerstone transformations of modern catalysis for the construction of carbon-carbon bonds. Despite its industrial and scientific centrality, most studies report yield as the sole performance metric, reducing an intrinsically dynamic process to a static end-point measurement. This simplification obscures the rich information embedded in concentration-time trajectories, which encode transient regimes, stiff dynamics, and the impact of ligands and base on catalytically relevant intermediates that are not experimentally observable. This thesis combines mechanistic simulation, hybrid deep learning for dynamical systems, and structural signal extraction from ligands to better understand the role of the catalytic system in governing reaction behavior. In the first part, ODE models derived from kinetic studies of increasing complexity are transformed into generative simulators to explore operational spaces via Latin Hypercube Sampling. The resulting in silico trajectories incorporate controlled noise, realistically reproducing experimental uncertainty. Sequential models (LSTMs) and continuous models constrained by chemical structure (stoichiometrically consistent Neural ODEs) are trained on these data, with loss functions restricted to experimentally measurable species. The results demonstrate that neural surrogates can learn latent catalytic dynamics and reproduce coherent behaviors even without supervision on palladium intermediates, outlining a new framework for hybrid modeling in catalysis. Subsequently, a retrospective extraction of structure-activity signals is performed using heterogeneous literature data, following rigorous curation of a reaction database containing catalyst-ligand complexes, in which ligands are converted into canonical SMILES. An Explainable Boosting Machine models yield as a function of operating conditions, isolating a residual signal corrected for these conditions. This signal is then subjected to motif mining using Morgan fingerprints and non-parametric statistical testing. Overall, the thesis proposes a reproducible computational platform that: (i) redefines kinetic modeling of the Suzuki-Miyaura reaction through structurally consistent and physically informed neural surrogates; and (ii) provides guidelines for the design of prospective datasets and catalysis studies oriented toward causal inference.| File | Dimensione | Formato | |
|---|---|---|---|
|
2026_03_Catalani_Executive_Summary.pdf
solo utenti autorizzati a partire dal 01/03/2029
Descrizione: Executive Summary
Dimensione
1.26 MB
Formato
Adobe PDF
|
1.26 MB | Adobe PDF | Visualizza/Apri |
|
2026_03_Catalani_Thesis.pdf
solo utenti autorizzati a partire dal 01/03/2029
Descrizione: Thesis
Dimensione
7.12 MB
Formato
Adobe PDF
|
7.12 MB | Adobe PDF | Visualizza/Apri |
I documenti in POLITesi sono protetti da copyright e tutti i diritti sono riservati, salvo diversa indicazione.
https://hdl.handle.net/10589/252312