Active Learning for Robust, High-Complexity Reactive Atomistic Simulations

Abstract

Machine learned reactive force fields based on polynomial expansions have been shown to be highly effective for describing simulations involving reactive materials. Nevertheless, the highly flexible nature of these models can give rise to a large number of candidate parameters for complicated systems. In these cases, reliable parameterization requires a well-formed training set, which can be difficult to achieve through standard iterative fitting methods. Here we present an active learning approach based on cluster analysis and Shannon information theory to enable semi-automated generation of informative training sets and robust machine learned force fields. Use of this tool is demonstrated for development of a model based on linear combinations of Chebyshev polynomials explicitly describing up to four-body interactions, for a chemically and structurally diverse system of C/O under extreme conditions. We show that this flexible training repository management approach enables development of models exhibiting excellent agreement with Kohn–Sham density functional theory (DFT) in terms of structure, dynamics, and speciation.

Using DFTB to Model Photocatalytic Anatase–Rutile TiO2 Nanocrystalline Interfaces and Their Band Alignment

“…This avoids reliance on iterative approaches that are required for nonlinear optimization problems (e.g., Levenberg–Marquardt) that are usually more computationally time-consuming and not guaranteed to result in the global minimum. ChIMES models have been extended to include four-body interactions in a similar fashion in reactive MD simulations, 16 , 19 though truncation of the ChIMES total energy with the three-body term has proven sufficient for determination of the DFTB repulsive energy. Note that DFTB in its original formulation uses a consistent two-center approximation in both the Hamiltonian matrix elements and the repulsive energy.…”

Calculation of the detonation state of HN3 with quantum accuracy

“…The development of the variable-charge ChIMES model is the subject of future work. ChIMES efforts are also currently underway to establish an active learning framework for automated selection/addition of isolated species 29 . The final training trajectory consisted of 60 original DFT frames (20 frames for each thermodynamic state), 60 frames from random displacement configurations, 420 gas molecules, and 22 frames taken from the iterative fitting scheme.…”

“…In its original formulation, ChIMES employs only two-and three-body interactions in the functional form 28 . Recently, it has been extended to also include four-body interactions 29 . The ChIMES model is parameterized by matching to reference data for selected configurations which are taken from short DFT-MD trajectories.…”

Development of Range-Corrected Deep Learning Potentials for Fast, Accurate Quantum Mechanical/molecular Mechanical Simulations of Chemical Reactions in Solution

“…18 In the past decade, researchers have developed a broad spectrum of different ML potentials. [19][20][21][22][23][24][25][26][27][28][29][30][31][32][33][34][35] Recently, an ML-based model called Deep Potential -Smooth Edition (DeepPot-SE) 36 was developed to efficiently represent organic molecules, metals, semiconductors and insulators with an accuracy comparable to that of ab initio QM models. The DeepPot-SE model has recently been highlighted in simulations of interfacial processes in aqueous aerosol 37 and large-scale combustion reactions in the gas phase, 38 and demonstrated great success in providing predictive insight into complex reaction processes.…”