Speaker: Markus Reiher, ETH Zürich

In this talk, I will discuss the latest developments of the Chemoton program [1] for automated reaction network explorations [2,3,4] from first principles. Chemoton is a versatile software for this purpose that we have been developing for almost a decade. It is freely available and open source as part of our SCINE package [5]. The capabilities of SCINE are broad and range from automated mechanism elucidation to rolling benchmarking with uncertainty quantification [6] and integrated microkinetic modeling [7,8,9].
Automated reaction network elucidation based on first-principles methods requires significant computational resources for the exploratory algorithms that crawl and search across Born-Oppenheimer potential energy surfaces. Accordingly, cheap quantum chemical methods such as density functional tight binding are often employed. However, they trade speed for accuracy with significant setbacks on the reliability of relative energies and even on network topology. Machine learning potentials (MLPs) promise to achieve comparatively high accuracy at the speed of force field evaluations. However, a significant initial training effort and the lack of MLP flexibility had hampered their application in an exploratory context. To address this problem, we developed lifelong-learning MLPs [10], which allow one to continually grow a knowledge base for quantum chemical reactivity studies. This new concept required significant methodological developments [11,12]. In view of the growing dominance of foundational models, which tend to become more and more important also for reaction exploration, lifelong data selection can be an important mechanism for fine-tuning such a general model to a specific reactivity study at hand.

References
[1] J. P. Unsleber, S. A. Grimmel, M. Reiher, Chemoton 2.0: Autonomous Exploration of Chemical Reaction Networks, J. Chem. Theory Comput. 2022, 18, 5393.
[2] G. N. Simm, A. C. Vaucher, M. Reiher, Exploration of Reaction Pathways and Chemical Transformation Networks, J. Phys. Chem. A, 2019, 123, 385.
[3] J. P. Unsleber, M. Reiher, The Exploration of Chemical Reaction Networks, Ann. Rev. Phys. Chem., 2020, 71, 121.
[4] M. Steiner, M. Reiher, Autonomous Reaction Network Exploration in Homogeneous and Heterogeneous Catalysis, Top. Catalysis, 2022, 65, 6.
[5] T. Weymuth, J. P. Unsleber, P. L. Tuertscher, M. Steiner, J.-G. Sobez, C. H. Mueller, M. M\"orchen, V. Klasovita, S. A. Grimmel, M. Eckhoff, K.-S. Csizi, F. Bosia, M. Bensberg, M. Reiher, SCINE-Software for chemical interaction networks, J. Chem. Phys. 2024, 160, 222501.
[6] G. Simm, M. Reiher, Error-Controlled Exploration of Chemical Reaction Networks with Gaussian Processes, J. Chem. Theory Comput. 2018, 14, 5238.
[7] J. Proppe, M. Reiher, Mechanism Deduction from Noisy Chemical Reaction Networks, J. Chem. Theory Comput. 2019, 15, 35.
[8] M. Bensberg, M. Reiher, Concentration-Flux-Steered Mechanism Exploration with an Organocatalysis Application, Isr. J. Chem. 2023, 63, e202200123.
[9] M. Bensberg, M. Reiher, Uncertainty-Aware First-Principles Exploration of Chemical Reaction Networks, J. Phys. Chem. A 2024, 128, 4532.
[10] M. Eckhoff, M. Reiher, Lifelong Machine Learning Potentials, J. Chem. Theory Comput. 2023, 19, 3509.
[11] M. Eckhoff, M. Reiher, CoRe optimizer: an all-in-one solution for machine learning, Mach. Learn.: Sci. Technol. 2024, 5, 015018.
[12] R. T. Husistein, M. Reiher, M. Eckhoff, NEAR: A Training-Free Pre-Estimator of Machine Learning Model Performance, ICLR, 2025, arXiv:2408.08776.