Tim Kircher

Microkinetic models provide physically grounded descriptions of catalytic systems. However, quantitative prediction of catalyst performance requires refinement of elementary-step parameters against reactor-level measurement data. This calibration is computationally demanding because it involves repeated reactor simulations with computationally expensive microkinetic models. Surrogates that approximate the kinetic model reduce evaluation cost, but are currently tied to fixed kinetic parameters and therefore cannot be used for mechanism calibration. To address this limitation, we introduce kinetically parameterized Global Reaction Neural Network (pGRNN) surrogates that map both operating conditions and elementary-step kinetic parameters of a detailed microkinetic mechanism to species source terms. By embedding stoichiometry, thermodynamic consistency, and elementary-step structure, the surrogate replaces expensive kinetics evaluations while retaining a mechanism-consistent representation. Using the preferential CO oxidation on platinum, we show that the pGRNN predicts source terms across a broad range of operating conditions and kinetic parameters with a relative error of 1.2 %. Reactor model evaluation is accelerated by a factor of 3100. The pGRNN is used in place of the kinetic model during mechanism calibration, reducing calibration time by factors of approximately 2400 for a synthetic single light-off curve and a synthetic multi-experiment campaign. Finally, we demonstrate the workflow by efficiently calibrating the mechanism against measured reactor-performance data from two experimental setups. A differentiable implementation of kinetic and reactor model further enables efficient gradient-based calibration. The proposed workflow provides an accelerated route for adapting detailed microkinetic models to catalyst-specific performance data.

Reaction kinetics lie at the center of reactive systems engineering, yet discovering them from data and incorporating them into predictive simulations remain time-consuming tasks that slow down model-based process design and control. To address this challenge, we introduce the Exponential Scaling Formation Consumption Neural Network (+/- ExpNet), a physics-inspired neural network architecture that can learn kinetics directly from integral reactor data and easily integrates with computational fluid dynamics to accelerate multiscale reactive simulations. The +/- ExpNet models latent formation and consumption rates which are subtracted to compute source terms, and enforces atom conservation as a hard constraint by exploiting the principles of elementary reaction kinetics. Using two catalytic and one non-catalytic test case we demonstrate that this relatively simple approach represents source terms with surprising accuracy, in most cases en par with previous approaches that require significantly more specific knowledge of the reaction system. Further, we employ the +/- ExpNet to extract kinetics for the direct DME synthesis using integral reactor data from literature. In doing so, we simultaneously derive kinetic models for both catalysts in this dual-catalyst system, without having to specify the global reactions occurring on the two catalysts or any thermodynamic constraints. This demonstrates that the +/- ExpNet provides a general and interpretable framework for automated kinetic modeling and efficient multiscale reactive simulations. Its simplicity makes it easy to integrate into big data frameworks and high-throughput experimentation, offering a path to make accurate kinetic models more widely accessible across catalysis, combustion, and other data-rich chemical disciplines.

Multiscale simulations of reactive flows are critical in many fields. However, their application is often hindered by the high computational cost of solving detailed chemical kinetics. Recent advances in surrogate models for reactive chemistry offer promising speedups, but ensuring physical consistency remains challenging. In particular, machine learning models for chemical kinetics must enforce atom balance and guarantee the positivity of predicted concentrations. Here, we introduce a positivity preserving projection and a correction by linear interpolation backtracking which simultaneously guarantee both constraints. We demonstrate this using two practical examples from atmospheric chemistry and heterogeneous catalysis, as well as for a large number of random, synthetically generated reaction systems. In all cases, our approach yields exclusively positive model predictions conforming to the atom balance, without reducing the overall accuracy of the model.

Multiscale simulations of reactive flows are critical in many fields. However, their application is often hindered by the high computational cost of solving detailed chemical kinetics. Recent advances in surrogate models for reactive chemistry offer promising speedups, but ensuring that these models remain physically consistent remains challenging. In particular, machine learning models for chemical kinetics must enforce atom balance and guarantee the positivity of predicted concentrations. Here, we introduce a positivity preserving projection that simultaneously enforces both constraints. We demonstrate this approach using an example from atmospheric chemistry and an example from heterogeneous catalysis. In both cases, the positivity preserving projection yields exclusively positive model predictions conforming to the atom balance, without reducing the overall accuracy of the model.

The digitalization of chemical research and industry generates large amounts of data, offering new opportunities for developing and parameterizing kinetic models. Leveraging this data requires machine learning techniques capable of autonomously extracting kinetics from reactor datasets. Recently, neural ordinary differential equations (neural ODEs) were coupled with reactor models to learn kinetics from ideal reaction systems, such as plug-flow reactors. However, real reactor set-ups commonly feature non-idealities such as heat- and mass transfer limitations, described by partial differential equations (PDEs) or differential-algebraic equations (DAEs). Discretizing such non-ideal PDE or DAE reactor models by finite volumes yields algebraic balance equations, which are solved by numerical schemes. In this work, we propose to learn kinetics from non-ideal reactor data using these balance equations and implicit neural networks, avoiding expensive backpropagation through the numerical solution by utilizing the implicit function theorem. The approach is demonstrated for the example of a mass transfer limited flat plate reactor for the preferential catalytic oxidation of CO. We show that Global Reaction Neural Networks, embedding thermodynamic and stoichiometric prior knowledge, coupled with a discretized two-phase CSTR cascade reactor model extract intrinsic kinetics from 50 integral reactor experiments. These kinetics generalize to new reactor geometries, where kinetics obtained by neural ODEs fail. Further, the approach is robust, recovering accurate kinetics even when training data is perturbed by 10% Gaussian noise. We expect that combining neural network-based kinetics with non-ideal reactor models will broaden the scope of kinetic model discovery and improve access to accurate kinetic models.