Abstract
We employ Gaussian process (GP) regression to adjust for systematic errors in D3-type dispersion corrections, resulting in a statistically improved model termed D3-GP. The model is trained on the differences between interaction energies obtained from PBE-D3(BJ)/ma-def2-QZVPP and high-level DLPNO-CCSD(T)/CBS calculations. Using a data set of 1,248 molecular dimers representing dispersion-dominated systems, we demonstrate that a reparametrized D3(BJ) model with a single empirical parameter ($a_2 = 5.6841$ bohr) outperforms existing parameter sets. To facilitate learning, we engineered two vectorial representations derived from the D3(BJ) interaction matrix: a distance-resolved interaction energy histogram (histD3(BJ)) and ordered eigenvalues (eigD3(BJ)). Furthermore, we introduce Batchwise Variance-Based Sampling (BVS), an active-learning scheme that utilizes GP prediction variance to select optimal training subsets. This approach yields a self-improving workflow capable of predicting system-focused, refined dispersion corrections with high efficiency.
Introduction
Density Functional Theory (DFT) is a cornerstone of modern computational chemistry, yet it fundamentally struggles to account for London dispersion interactions. These weak, long-range attractive forces are critical for the stability and structure of many molecular systems. To address this, empirical "patches" like the D3(BJ) dispersion correction are commonly applied. While computationally efficient and physically grounded, these corrections often suffer from inherent model inadequacy, failing to perfectly replicate high-level quantum chemical results. The goal of this work is to bridge the gap between cheap, approximate methods and "gold standard" reference calculations like CCSD(T) by using machine learning to learn and correct the systematic errors of the D3 model.
Methods
The D3-GP approach treats the error of the D3 model as a residual to be learned via Gaussian Process regression, a technique known as $\Delta$-learning.
Reparametrization of D3(BJ)
Analysis of dispersion-dominated systems revealed that the standard D3(BJ) model might be over-parameterized for specific chemical domains. By setting the empirical parameters $a_1$ and $s_8$ to zero and optimizing the remaining parameter $a_2$, we achieved superior performance for the target systems compared to universal parameter sets. This simplified model serves as the baseline for further GP refinement.
Feature Engineering
To represent molecular systems without the complexities of raw atomic coordinates, we developed two descriptors based directly on the D3 interaction matrix:
- histD3(BJ): A histogram of atom-pairwise interaction energies, resolved by distance.
- eigD3(BJ): The eigenvalues of the interaction matrix, ordered by absolute value. These descriptors ensure the model remains "physics-informed" by mapping directly from D3 information, making the correction comparable to the original approach.
Batchwise Variance-Based Sampling (BVS)
Active learning is central to the D3-GP workflow. By leveraging the uncertainty (variance) provided by the GP, the BVS algorithm identifies batches of molecular configurations where the model is most uncertain. These points are then prioritized for expensive reference calculations, maximizing the information gain per calculation and allowing the model to "self-improve" efficiently.
Data Sets
The study utilized a comprehensive collection of 1,248 molecular dimers, categorized into three primary sets to test different aspects of the model's performance, including equilibrium structures and various relative orientations.
The S13x8 set provides a benchmark for dispersion-dominated systems at various distances, while the ROTA and CONF sets explore the effects of relative orientation and internal molecular conformation of ethyne–pentane dimers, respectively.
Key Results and Discussion
The D3-GP model demonstrates significant improvements over standard dispersion corrections. The reparametrized D3 model alone reduced the Mean Absolute Relative Error (MARE) on the S13x8-V set from 13.4% (using standard Grimme parameters) to 8.2%. When combined with GP refinement, the model exhibits a steep learning curve, often outperforming standard D3 after seeing fewer than five training examples for specific systems.
The BVS strategy proved highly efficient, reaching accuracy thresholds (e.g., 0.05 kcal/mol) with far fewer data points than random sampling. Among the descriptors, histD3(BJ) generally provided more stable learning and lower prediction errors. However, challenges remain in extrapolation; while the model excels at interpolation within its training domain, its performance on chemically diverse systems outside that domain shows more modest improvements, suggesting that training set selection is critical.
Conclusion
D3-GP represents a shift toward self-improving computational tools in quantum chemistry. By integrating physics-based models with active-learning-driven machine learning, we can achieve high-level accuracy at a fraction of the cost of traditional "gold standard" methods. Future developments will focus on expanding the model's domain to include more diverse interactions, such as hydrogen bonding, and establishing a community-driven database for continuous model refinement.
Summary updated