An Intelligent System for Reaction Kinetic Modeling and Catalyst Design

Extracting kinetic information in catalysis: an automated tool for the exploration of small data

Kinetically relevant information for heterogeneously catalysed reactions is automatically extracted from small datasets by means of a newly-developed machine learning chemically-enriched tool.

Learning Chemistry of Complex Reaction Systems via a Python First-Principles Reaction Rule Stencil (pReSt) Generator

Complex reaction networks can be generated with automated network generators from initial reactants and reaction rules. Reaction rule specification is central to network generation. These reaction rules are, at present, user-defined based on (intuitive) expert knowledge of chemistry and are often transferred from gas-phase to surface processes. The catalyst active site geometry is usually left out but is often responsible for selectivity. We propose a first-principles-based reaction mechanism generation framework using density functional theory (DFT) data of published reaction mechanisms. The framework "learns the chemistry" from published mechanisms. It can generate reaction networks not studied before, "flag" reactions not seen before for further DFT convergence tests, and easily reconcile differences between catalysts and reactants that may introduce new pathways never seen before. As such, it can be a diagnostic tool for data (mechanism) quality assessment and novel pathway discovery to new molecules. A software, the Python Reaction Stencil (pReSt), was developed for this purpose to wrap around automatic mechanism generation software. Multiple catalytic chemistries are considered to show the efficacy of the proposed framework.

Identification of Granule Growth Regimes in High Shear Wet Granulation Processes Using a Physics-Constrained Neural Network

The digitization of manufacturing processes has led to an increase in the availability of process data, which has enabled the use of data-driven models to predict the outcomes of these manufacturing processes. Data-driven models are instantaneous in simulate and can provide real-time predictions but lack any governing physics within their framework. When process data deviates from original conditions, the predictions from these models may not agree with physical boundaries. In such cases, the use of first-principle-based models to predict process outcomes have proven to be effective but computationally inefficient and cannot be solved in real time. Thus, there remains a need to develop efficient data-driven models with a physical understanding about the process. In this work, we have demonstrate the addition of physics-based boundary conditions constraints to a neural network to improve its predictability for granule density and granule size distribution (GSD) for a high shear granulation process. The physics-constrained neural network (PCNN) was better at predicting granule growth regimes when compared to other neural networks with no physical constraints. When input data that violated physics-based boundaries was provided, the PCNN identified these points more accurately compared to other non-physics constrained neural networks, with an error of <1%. A sensitivity analysis of the PCNN to the input variables was also performed to understand individual effects on the final outputs.

Computational design of heterogeneous catalysts and gas separation materials for advanced chemical processing

Functional materials are widely used in chemical industry in order to reduce the process cost while simultaneously increase the product quality. Considering their significant effects, systematic methods for the optimal selection and design of materials are essential. The conventional synthesis-and-test method for materials development is inefficient and costly. Additionally, the performance of the resulting materials is usually limited by the designer’s expertise. During the past few decades, computational methods have been significantly developed and they now become a very important tool for the optimal design of functional materials for various chemical processes. This article selectively focuses on two important process functional materials, namely heterogeneous catalyst and gas separation agent. Theoretical methods and representative works for computational screening and design of these materials are reviewed.

A quantitative uncertainty metric controls error in neural network-driven chemical discovery

Machine learning (ML) models, such as artificial neural networks, have emerged as a complement to high-throughput screening, enabling characterization of new compounds in seconds instead of hours. The promise of ML models to enable large-scale chemical space exploration can only be realized if it is straightforward to identify when molecules and materials are outside the model's domain of applicability. Established uncertainty metrics for neural network models are either costly to obtain (e.g., ensemble models) or rely on feature engineering (e.g., feature space distances), and each has limitations in estimating prediction errors for chemical space exploration. We introduce the distance to available data in the latent space of a neural network ML model as a low-cost, quantitative uncertainty metric that works for both inorganic and organic chemistry. The calibrated performance of this approach exceeds widely used uncertainty metrics and is readily applied to models of increasing complexity at no additional cost. Tightening latent distance cutoffs systematically drives down predicted model errors below training errors, thus enabling predictive error control in chemical discovery or identification of useful data points for active learning.

Descriptor–property relationships in heterogeneous catalysis: exploiting synergies between statistics and fundamental kinetic modelling

Combined kinetic and statistical approach to shed light on the link between kinetically-relevant descriptors and easily tuneable catalyst properties.

Reduced-order modelling of biochemical networks: application to the GTPase-cycle signalling module

Biochemical systems embed complex networks and hence development and analysis of their detailed models pose a challenge for computation. Coarse-grained biochemical models, called reduced-order models (ROMs), consisting of essential biochemical mechanisms are more useful for computational analysis and for studying important features of a biochemical network. The authors present a novel method to model-reduction by identifying potentially important parameters using multidimensional sensitivity analysis. A ROM is generated for the GTPase-cycle module of m1 muscarinic acetylcholine receptor, Gq, and regulator of G-protein signalling 4 (a GTPase-activating protein or GAP) starting from a detailed model of 48 reactions. The resulting ROM has only 17 reactions. The ROM suggested that complexes of G-protein coupled receptor (GPCR) and GAP--which were proposed in the detailed model as a hypothesis--are required to fit the experimental data. Models previously published in the literature are also simulated and compared with the ROM. Through this comparison, a minimal ROM, that also requires complexes of GPCR and GAP, with just 15 parameters is generated. The proposed reduced-order modelling methodology is scalable to larger networks and provides a general framework for the reduction of models of biochemical systems.