Döppel, Felix Antonidas (2024)
Physics-Enhanced Machine Learning for Chemical Kinetics.
Technische Universität Darmstadt
doi: 10.26083/tuprints-00027384
Dissertation, Erstveröffentlichung, Verlagsversion
Kurzbeschreibung (Abstract)
The energy transition and the transformation of the chemical industry are major efforts in addressing the challenges of climate change. Both require the development of new and optimized catalytic devices. The rational design of these devices depends on a thorough understanding of the underlying chemical kinetics. As kinetic model development gets outpaced by the ever-increasing availability of kinetic data, there is a growing demand for automated mechanism discovery. But even when detailed kinetic models are available, their use for the knowledge-based design of catalytic devices is limited by their high computational cost. The objective of this work is to design physically plausible machine learning models that facilitate the discovery of chemical reaction mechanisms and enable their efficient implementation in simulations of catalytic devices. The cost of reactive flow simulations is often reduced by replacing the computationally intensive evaluation of chemical kinetics with a numerically less demanding, so-called surrogate model. However, surrogates have not yet been systematically applied to surface kinetics of industrially relevant complexity because they rely on logarithmic data normalization. The logarithm, however, is incompatible with systems that contain intermediate species or operate close to the chemical equilibrium, as the modelled source terms change sign and thereby leave the domain of the logarithm. This work provides two methods that extend the scope of surface kinetic surrogate modelling: 1. Modelling the rates of the rate-determining steps instead of species source terms re-enables logarithmic normalization because the considered elementary rates are strictly positive. 2. Replacing the logarithm with specialized logarithm-like functions extends the domain of allowed source terms to positive as well as negative numbers. Further, this work introduces neural networks to surface kinetic modelling and demonstrates their superiority over splines, the former state of the art, in terms of accuracy, prediction time, and the required amount of storage space. The proposed latent data transformation technique makes use of the high structural flexibility of neural network models to embed the principles of atom conservation, the Arrhenius law, and the law of mass action directly into the model. The resulting surrogates accurately predict the chemical kinetics of industrially relevant systems, as exemplarily shown for the preferential oxidation of CO, which is relevant to hydrogen production for fuel cell applications, and the ammonia oxidation under industrially relevant conditions of the Ostwald process. Reactor simulations of these systems are accelerated by four to five orders of magnitude. Increasingly detailed kinetics, such as first principles kinetic Monte Carlo simulations, are becoming the gold standard in chemical engineering. Their solution is computationally so expensive that even the collection of a sufficient amount of data to train a surrogate model becomes infeasible. This work proposes a grid-free training set design scheme that evaluates only those data that significantly contribute to the accuracy of the surrogate. Applied to the preferential oxidation of CO, this leads to a 14-fold reduction in the amount of required training data. Uncertainty estimation is performed by two-layered kernel models and further employed to track the accuracy of surrogates during their use in reactor simulations. This allows to refine the model on-the-fly and thereby ensures reliable simulation results, even outside the original training range of the model. The recently developed chemical reaction neural network is a digital twin of the classic microkinetic mechanism that has found widespread application in many fields, such as (bio-)chemical engineering and combustion. While it encodes some fundamental physical laws, mass and atom conservation are still violated. Here, atom conservation is enforced by adding a dedicated neural network layer which can be interpreted as constraining the model to physically realizable stoichiometries. As the resulting models are physically consistent, they are more robust to limited data availability, noisy data, and systematic measurement errors. Overall, this work improves the physical interpretability and extrapolation capabilities of machine learning models for chemical kinetics. In particular, it presents physics-enhanced neural network architectures, that contain the fundamental physical laws of atom conservation and thermodynamics as well as the more specific Arrhenius law and the law of mass action. Together with accurate uncertainty quantification, this yields fast and reliable implementations of chemical kinetics into reactive flow simulations, allowing their systematic evaluation for the knowledge-based design of catalytic devices. Further, these models perform automated mechanism discovery, pushing the boundary of kinetic insights.
Typ des Eintrags: | Dissertation | ||||
---|---|---|---|---|---|
Erschienen: | 2024 | ||||
Autor(en): | Döppel, Felix Antonidas | ||||
Art des Eintrags: | Erstveröffentlichung | ||||
Titel: | Physics-Enhanced Machine Learning for Chemical Kinetics | ||||
Sprache: | Englisch | ||||
Referenten: | Votsmeier, Prof. Dr. Martin ; Weeger, Prof. Dr. Oliver ; Deutschmann, Prof. Dr. Olaf | ||||
Publikationsjahr: | 24 Mai 2024 | ||||
Ort: | Darmstadt | ||||
Kollation: | IX, 110 Seiten | ||||
Datum der mündlichen Prüfung: | 13 Mai 2024 | ||||
DOI: | 10.26083/tuprints-00027384 | ||||
URL / URN: | https://tuprints.ulb.tu-darmstadt.de/27384 | ||||
Kurzbeschreibung (Abstract): | The energy transition and the transformation of the chemical industry are major efforts in addressing the challenges of climate change. Both require the development of new and optimized catalytic devices. The rational design of these devices depends on a thorough understanding of the underlying chemical kinetics. As kinetic model development gets outpaced by the ever-increasing availability of kinetic data, there is a growing demand for automated mechanism discovery. But even when detailed kinetic models are available, their use for the knowledge-based design of catalytic devices is limited by their high computational cost. The objective of this work is to design physically plausible machine learning models that facilitate the discovery of chemical reaction mechanisms and enable their efficient implementation in simulations of catalytic devices. The cost of reactive flow simulations is often reduced by replacing the computationally intensive evaluation of chemical kinetics with a numerically less demanding, so-called surrogate model. However, surrogates have not yet been systematically applied to surface kinetics of industrially relevant complexity because they rely on logarithmic data normalization. The logarithm, however, is incompatible with systems that contain intermediate species or operate close to the chemical equilibrium, as the modelled source terms change sign and thereby leave the domain of the logarithm. This work provides two methods that extend the scope of surface kinetic surrogate modelling: 1. Modelling the rates of the rate-determining steps instead of species source terms re-enables logarithmic normalization because the considered elementary rates are strictly positive. 2. Replacing the logarithm with specialized logarithm-like functions extends the domain of allowed source terms to positive as well as negative numbers. Further, this work introduces neural networks to surface kinetic modelling and demonstrates their superiority over splines, the former state of the art, in terms of accuracy, prediction time, and the required amount of storage space. The proposed latent data transformation technique makes use of the high structural flexibility of neural network models to embed the principles of atom conservation, the Arrhenius law, and the law of mass action directly into the model. The resulting surrogates accurately predict the chemical kinetics of industrially relevant systems, as exemplarily shown for the preferential oxidation of CO, which is relevant to hydrogen production for fuel cell applications, and the ammonia oxidation under industrially relevant conditions of the Ostwald process. Reactor simulations of these systems are accelerated by four to five orders of magnitude. Increasingly detailed kinetics, such as first principles kinetic Monte Carlo simulations, are becoming the gold standard in chemical engineering. Their solution is computationally so expensive that even the collection of a sufficient amount of data to train a surrogate model becomes infeasible. This work proposes a grid-free training set design scheme that evaluates only those data that significantly contribute to the accuracy of the surrogate. Applied to the preferential oxidation of CO, this leads to a 14-fold reduction in the amount of required training data. Uncertainty estimation is performed by two-layered kernel models and further employed to track the accuracy of surrogates during their use in reactor simulations. This allows to refine the model on-the-fly and thereby ensures reliable simulation results, even outside the original training range of the model. The recently developed chemical reaction neural network is a digital twin of the classic microkinetic mechanism that has found widespread application in many fields, such as (bio-)chemical engineering and combustion. While it encodes some fundamental physical laws, mass and atom conservation are still violated. Here, atom conservation is enforced by adding a dedicated neural network layer which can be interpreted as constraining the model to physically realizable stoichiometries. As the resulting models are physically consistent, they are more robust to limited data availability, noisy data, and systematic measurement errors. Overall, this work improves the physical interpretability and extrapolation capabilities of machine learning models for chemical kinetics. In particular, it presents physics-enhanced neural network architectures, that contain the fundamental physical laws of atom conservation and thermodynamics as well as the more specific Arrhenius law and the law of mass action. Together with accurate uncertainty quantification, this yields fast and reliable implementations of chemical kinetics into reactive flow simulations, allowing their systematic evaluation for the knowledge-based design of catalytic devices. Further, these models perform automated mechanism discovery, pushing the boundary of kinetic insights. |
||||
Alternatives oder übersetztes Abstract: |
|
||||
Freie Schlagworte: | Machine Learning, Chemical Kinetics, Physics-Informed Neural Networks, Reactor Simulation, Surrogate Modeling, Atom Balance, Surface Kinetics, Rate-determining Step, Uncertainty Quantification, Latent Space, Kernel Models, Mechanism Discovery, Chemical Reaction Neural Network, Neural ODE, Null Space, Key Species | ||||
Status: | Verlagsversion | ||||
URN: | urn:nbn:de:tuda-tuprints-273848 | ||||
Sachgruppe der Dewey Dezimalklassifikatin (DDC): | 000 Allgemeines, Informatik, Informationswissenschaft > 004 Informatik 500 Naturwissenschaften und Mathematik > 540 Chemie 600 Technik, Medizin, angewandte Wissenschaften > 660 Technische Chemie |
||||
Fachbereich(e)/-gebiet(e): | 07 Fachbereich Chemie 07 Fachbereich Chemie > Ernst-Berl-Institut 07 Fachbereich Chemie > Ernst-Berl-Institut > Fachgebiet Technische Chemie |
||||
TU-Projekte: | Bund|05M20RDA|ML-MORE | ||||
Hinterlegungsdatum: | 24 Mai 2024 12:01 | ||||
Letzte Änderung: | 27 Mai 2024 06:48 | ||||
PPN: | |||||
Referenten: | Votsmeier, Prof. Dr. Martin ; Weeger, Prof. Dr. Oliver ; Deutschmann, Prof. Dr. Olaf | ||||
Datum der mündlichen Prüfung / Verteidigung / mdl. Prüfung: | 13 Mai 2024 | ||||
Export: | |||||
Suche nach Titel in: | TUfind oder in Google |
![]() |
Frage zum Eintrag |
Optionen (nur für Redakteure)
![]() |
Redaktionelle Details anzeigen |