Files and links (1)
Jonah_Marks_Thesis-45.66 MB
Embargoed Access, Embargo ends: 01/23/2027
Dissertation
An improved freezing string method for fast and reliable transition state searches with machine learning interatomic potentials
University of Iowa
Doctor of Philosophy (PhD), University of Iowa
Autumn 2025
Abstract
The reliable and efficient discovery of transition states remains one of the central challenges in theoretical chemistry, underpinning our ability to predict reactivity, design catalysts, and model chemical mechanisms across complex systems. Conventional ab initio transition state search methods are computationally demanding, limiting their application to small systems and well-behaved potential energy surfaces. This thesis develops a framework that combines advances in reaction-path algorithms with modern machine-learned interatomic potentials (MLIPs) to enable accurate, automated, and scalable transition state searches at a fraction of the traditional computational cost.
In the first part of this work, an improved variant of the Freezing String Method (FSM) is developed that incorporates interpolation in redundant internal coordinates and advanced optimization algorithms. Our results show that incorporation of internal coordinates interpolation improves the reliability of the FSM, enabling larger interpolation step sizes and fewer optimization steps per cycle, which together yield nearly a 50% reduction in computational cost associated with required density functional theory (DFT) calculations while maintaining a 100% success rate on benchmark chemical reaction test cases, including systems where previous attempts based on linear synchronous transit interpolation have failed. We provide an open-source Python implementation of the FSM, in addition to the reactant, product, and transition state structures of all reactions studied.
The second component of this work integrates graph neural network (GNN) potential energy surfaces with the FSM to reduce dependence on repeated ab initio calculations. A SchNet GNN model pre-trained on off-equilibrium structures of small organic molecules from the ANI-1
quantum chemical dataset and fine-tuned using the GDB7-20-TS dataset—comprising reactant, product, and transition state structures— accurately reproduces transition state regions, enabling FSM-based searches with a 100% success rate and a 72% reduction in DFT calculations. This
study establishes that with appropriate datasets and training, GNN potentials can serve as reliable surrogates for DFT, maintaining chemical accuracy while dramatically improving efficiency.
The final part of this thesis presents a systematic benchmark of hybrid transition state search workflows combining six freely available potentials (MACE-OMol25, UMA-Small, UMA-Medium, eSEN-S, AIMNet2, and GFN2-xTB) with two reaction path-finding algorithms (FSM and climbing-image nudged elastic band) across 58 diverse reactions spanning small organics, polymerization chemistry, and transition-metal catalysis. We find that models trained on the chemically diverse Open Molecules 2025 dataset exhibit markedly superior performance, with MACE-OMol25 achieving a 96.6% success rate while requiring fewer than four DFT gradient evaluations per reaction on organic systems—a 94-96% reduction compared to conventional DFT-based searches. Low-level refinement on the MLIP surface before high-level DFT optimization reduces computational cost three-fold with minimal loss in reliability. For transition-metal systems, UMA-Medium demonstrates promising transferability to in-distribution transition metal complex reactions and out-of-distribution heterogeneous catalysis test cases. These results establish MLIP accelerated workflows as practical tools for automated reaction discovery, enabling near-ab initio accuracy at a fraction of traditional expense.
Collectively, this thesis demonstrates that coupling efficient reaction path-finding algorithms with modern MLIPs enables near-ab initio accuracy at over 95% lower computational cost. The resulting ML-FSM framework provides a fast, open-source, and reproducible foundation for high-throughput exploration of chemical reactivity, catalysis, and materials design.
Details
- Title: Subtitle
- An improved freezing string method for fast and reliable transition state searches with machine learning interatomic potentials
- Creators
- Jonah Marks
- Contributors
- Joseph Gomes (Advisor)Charles Stanier (Committee Member)Sara Mason (Committee Member)David Rethwisch (Committee Member)Mike Schnieders (Committee Member)Chris Coretsopoulos (Committee Member)
- Resource Type
- Dissertation
- Degree Awarded
- Doctor of Philosophy (PhD), University of Iowa
- Degree in
- Chemical and Biochemical Engineering
- Date degree season
- Autumn 2025
- Publisher
- University of Iowa
- Number of pages
- xv, 133 pages
- Copyright
- Copyright 2025 Jonah Marks
- Language
- English
- Date submitted
- 12/08/2025
- Description illustrations
- illustrations (some color)
- Description bibliographic
- Includes bibliographical references.
- Public Abstract (ETD)
Chemical reactions are foundational to many critical processes of the modern world, from creating new medicines and materials to understanding how pollutants break down in the atmosphere. To study how these reactions happen, scientists often rely on computer simulations that predict how atoms move and bonds form or break. The most accurate methods for doing this, based on quantum mechanics, can take days or even weeks of computer time for a single reaction, making it difficult to explore new ideas quickly.
This dissertation develops new ways to make those simulations dramatically faster without losing accuracy. It introduces algorithms that find the transition state, the fleeting atomic arrangement that represents the highest-energy point along a reaction, using advanced mathematical strategies to trace reaction pathways efficiently. These methods are then combined with artificial-intelligence models trained on large datasets of quantum-chemical calculations, allowing a computer to approximate the behavior of electrons almost instantly.
Together, these developments make it possible to study chemical reactions hundreds of times faster than before, enabling rapid discovery of catalysts, materials, and chemical processes that were previously too expensive to simulate. The software created through this work is opensource, and already being adopted by researchers and industry, helping bring powerful computational chemistry tools into routine scientific and technological use.
- Academic Unit
- Chemical and Biochemical Engineering
- Record Identifier
- 9985134948802771
Metrics
1 Record Views