Jason D. Goodpaster

Chemical reactions in solution are central to biological function, synthetic chemistry, and materials design. Accurate modeling of these systems is essential for obtaining mechanistic insights but remains computationally demanding. Hybrid machine-learned/molecular mechanics (ML/MM) simulations offer a promising compromise between quantum accuracy and computational efficiency. However, most existing ML/MM methods exclude solvent molecules from the ML region, limiting their ability to capture solvent-mediated reactivity. In this study, we introduce a solvent-inclusive ML/MM methodology that addresses this limitation. Our approach defines the ML and MM regions using a fixed spatial boundary, avoiding costly topology updates, and handles interactions across the boundary using a force-partitioning scheme. We evaluate our framework through simulations of bulk water and the solvent-mediated dissociation of formic acid using different ML region sizes. Our results reveal that structural and dynamical properties of bulk water are preserved for sufficiently large ML regions. Notably, we show that the choice of the potential energy surface representation for the ML and MM regions can affect these properties, especially at smaller ML region sizes. Our calculations of the free energy profiles for formic acid dissociation also show trends consistent with reference systems, but deviations are observed. We comment on potential reasons for these observations. Overall, our study highlights the promise of ML/MM approaches that are solvent-inclusive and provides directions for further development.

Chemical reactions in solution are central to biological function, synthetic chemistry, and materials design. Accurate modeling of these systems is essential for obtaining mechanistic insights, but remains computationally demanding. Hybrid machine-learned/molecular mechanics (ML/MM) simulations offer a promising compromise between quantum accuracy and computational efficiency. However, most existing ML/MM methods exclude solvent molecules from the ML region, limiting their ability to capture solvent-mediated reactivity. In this study, we introduce a solvent-inclusive ML/MM methodology that addresses this limitation. Our approach defines the ML and MM regions using a fixed spatial boundary, avoiding costly topology updates, and handles interactions across the boundary using a force-partitioning scheme. We evaluate our framework through simulations of bulk water and the solvent-mediated dissociation of formic acid using different ML region sizes. Our results reveal that structural and dynamical properties of bulk water are preserved for sufficiently large ML regions. Notably, we show that the choice of the potential energy surface representation for the ML and MM regions can affect these properties, especially at smaller ML region sizes. Our calculations of the free energy profiles for formic acid dissociation also show trends consistent with reference systems, but deviations are observed. We comment on potential reasons for these observations. Overall, our study highlights the promise of ML/MM approaches that are solvent-inclusive and provides directions for further development.

We study the gap energy of the semiconducting oxide SnO2 through ab initio calculations including both density functional theory (DFT) and coupled cluster methods. The effectiveness of twist averaging in reducing finite-size errors is evaluated across different functionals. We report an overestimation of gap energy when applying finite-size scaling to reach the thermodynamic limit in equation-of-motion (EOM) CCSD calculations. To mitigate one-body and many-body errors, we integrate twist averaging with a post-processing correction mechanism that compares finite-size and infinite-size DFT calculations using hybrid functionals. While inspired by the Kwee, Zhang, and Krakauer approach, our method is specifically tailored to hybrid functionals for a more accurate treatment of exchange-correlation effects. Our approach ensures that the many-body interactions are accurately captured in the estimated gap for an infinite system. We introduce unique single twist angles that provide cost-effective and accurate energies compared to to full twist averaging in EOM-CCSD calculations. Applying this approach to SnO2, we calculate a fundamental gap of 3.46 eV, which closely matches the 3.59 eV gap obtained from two-photon spectroscopy experiments, demonstrating the accuracy of this method.

We study the gap energy of the semiconducting oxide SnO2 through ab-initio calculations including both DFT and coupled cluster methods. The effectiveness of twist-averaging in reducing finite-size errors is evaluated across different functionals. We report an overestimation of gap energy when applying finite-size scaling at the thermodynamic limit in equation-of-motion (EOM) CCSD calculations. To mitigate one-body and many-body errors, we integrate twist averaging with a post-processing correction mechanism, comparing finite-size and infinite-size DFT calculations using hybrid functionals. While inspired by the Kwee, Zhang, and Krakauer (KZK) approach, our method is tailored to hybrid functionals for a more accurate treatment of exchange-correlation effects. Our approach ensures that the many-body interactions are accurately reflected in the estimated gap for an infinite system. We introduce unique single twist angles that yield cost- effective and accurate energies, in comparison to full twist averaging in the EOM-CCSD calculations. Applying this approach to SnO2, we calculate a fundamental gap of 3.46 eV, closely matching the 3.59 eV gap obtained from two-photon spectroscopy experiments, thereby demonstrating the accuracy of this method.

Machine-learning models for predicting adsorption energies on metallic surfaces often rely on basic elemental properties and electronic and geometric descriptors. Here, we apply categorical entity embedding, a featurization method inspired by natural language processing techniques, to predict adsorption energies on bimetallic alloy surfaces using categorical descriptors. Using this method, we develop a machine-learned representation from categorical descriptors (e.g., surface composition, adsorbate type, and site type) of the slab/adsorbate complex. By combining this representation with numerical features (e.g., slab metal stoichiometric ratios), we create the CatEmbed representation. Remarkably, decision tree models trained using CatEmbed, which includes no explicit geometric information, achieve a Mean Absolute Error (MAE) of 0.12 eV. Additionally, we extend this technique to predict reaction energies on bimetallic surfaces, creating the CatEmbed-React representation, which achieves an MAE of 0.08 eV. These findings highlight the effectiveness of categorical entity embedding for predicting adsorption and reaction energies on bimetallic alloy surfaces.

Machine-learning models for predicting adsorption energies on metallic surfaces often relies on basic elemental properties, electronic, and geometric descriptors. Here, we apply categorical entity embedding, a featurization method inspired by natural language processing techniques, to predict adsorption energies on bimetallic alloy surfaces using categorical descriptors. Using this method, we develop a machine-learned representation from categorical descriptors (e.g., surface composition, adsorbate type, and site type) of the slab/adsorbate complex. By combining this representation with numerical features (e.g., slab metal stoichiometric ratios), we create the CatEmbed representation. Remarkably, decision tree models trained using CatEmbed, which includes no explicit geometric information, achieve a Mean Absolute Error (MAE) of 0.12 eV. Additionally, we extended this technique to predict reaction energies on bimetallic surfaces, creating the CatEmbed-React representation, which achieves an MAE of 0.08 eV. These findings highlight the effectiveness of categorical entity embedding for predicting adsorption and reaction energies on bimetallic alloy surfaces.

Understanding the properties of the adsorbed oxygen species on the SnO2 surface is essential in describing the reaction mechanisms responsible for gas sensing applications. Density functional theory (DFT) calculations were performed to shed light on these mechanisms by exploring the energetics of molecular, atomic, and dissociative oxygen adsorption on various types of reduced, defective, and stoichiometric SnO2(110) surfaces. Charge analysis distinguished various forms of charged oxygen species on the surface. The calculations identified the existence of O_2^(2-), O_2^-, and O^(2-) on the reduced surface and O_2^(2-), O^-, and O^(2-) on the defective and the oxidized surfaces. The role of oxygen vacancies in recognizing the adsorption modes and relevant properties is demonstrated. The described electronic structure in this work supports experimental findings.

Large, condensed phased, and extended systems remain a challenge for theoretical studies due to the compromise between accuracy and computational cost in their calculations. Machine learning methods are on the rise to solve this trade off by training on large datasets of highly accurate calculations that are traditionally hard to obtain. The development of interatomic machine learning potentials has resulted in the ability to model high-quality potential energy surfaces with near ab initio level of accuracy at low computational cost. However, just like other machine learning applications, such methods face challenges when it comes to quality training data and transferability, specifically to systems of chemical space beyond its training. In this work, we present the continuous exploration of utilizing machine learning methods to build and achieve accurate and efficient potential energy surface for bond dissociation and reactive chemistry, and explore sampling techniques that can allow interatomic neural network potentials designed to model potential energy surfaces, such as ANI and NequIP, to accurately predict bond dissociation energy and model reactive chemistry, and to obtain transferability beyond its training data across chemical space.

Neural network potentials achieve CASPT2 accuracy for reactive chemistry and molecular simulations. Using transfer learning, these potentials require minimal CASPT2 data on small systems to accurately predict bond dissociation in larger systems.

We employ an active learning scheme to train a neural network potential which is then utilized in RPMD simulations to calculate the thermal rate coefficients for a hydrogen atom abstraction from methane by hydroxyl radical. We use a active learning approach to determine which geometries from umbrella sampling trajectories, performed using our neural network potential, should be included in our CCSD(T) training data to improve the neural network potential. After sufficient accuracy of the neural network potential is achieved, we use this potential energy surface in RPMD trajectories to calculate the reaction rate using the Bennet-Chandler method. We find this approach to be exceptionally accurate, with errors in the rate less than 20% at 200K and decreasing to 10% at 1000K. Next, we examine the limit of how much data was required to obtain highly accurate neural network potentials and find that only 7,000 CCSD(T) are required to preserve the accuracy of the network. Finally, we examine a transfer learning approach for this reaction, and find that transfer learning does not provide sufficient accuracy for rate calculations.