Active Learning
Active learning is an iterative strategy in which a model trained on the data collected so far decides which experiments or calculations to run next, typically by trading off exploration of uncertain regions against exploitation of important ones. It aims to reach a given level of knowledge with fewer samples than designs fixed in advance, and Bayesian optimization is a common special case when the goal is to optimize a property.
Two-network disagreement to grow a potential’s training set (2018)
To build a Pd neural network potential, Gao and Kitchin trained two networks of different architecture, ran MD or other simulations with one, and sent to DFT the configurations where the two predicted very different energies. Adding these structures and retraining improved the potential systematically and self-consistently, so that 3035 targeted calculations sufficed for bulk, defect and surface properties, rather than the tens of thousands often used for general-purpose potentials (Gao 2018). Summary: gao2018.
Neural network ensemble active learning for batch relaxations (2021)
Yang, Jiménez-Negrón and Kitchin used a 10-member SingleNN ensemble as a surrogate for DFT during geometry optimization, sending a structure to DFT whenever its ensemble energy variance exceeded twice the largest variance on the training set (Yang et al. 2021). Relaxing many related configurations with one shared ensemble beat single-configuration active learning, and pretraining on prior trajectories helped further: for 13 acrolein/AgPd(111) configurations needing about 193 quasi-Newton DFT steps each, calls dropped about 11-fold (single), 13-fold (multiple) and more with warm-up, versus about 4-fold for a GP. Climbing-image NEB runs needed 9 instead of 596 EMT calls and 30 instead of 1109 DFT calls. Summary: yang2021.
Delta method uncertainty to select retraining data (2022)
Zhan and Kitchin used delta method standard errors to detect where an AuPd neural network potential, trained at one lattice constant, was extrapolating to two others, then added 365 configurations from each new set and retrained (Zhan 2021). After retraining the predictions returned to parity, interval coverage was about 98.5%, and the standard errors on the new sets dropped to the level of the training set, illustrating an uncertainty-driven loop that could also run on the fly during molecular dynamics. They note the Fisher information can be approximately updated by adding the new points’ contributions when parameters change little. Summary: zhan2022.
Acquisition rules for classification boundaries (2023)
Bhat and Kitchin adapted active learning to binary classification, where the goal is a boundary rather than an optimum (Bhat 2023). Using a support vector classifier with an RBF kernel and its class probabilities, they designed three acquisition rules (broad initial sampling, low point density, and path tracing) that place batches of new samples near the current predicted boundary, and they scored runs on number of experiments, overlap with the true region, sample distance from the boundary and 95% confidence band area. Path tracing had the best combined score in all three test problems, with broad initial sampling close behind, while the more exploratory low point density rule cost the most experiments because of over-extrapolation. The methods are distributed in the boundarysearch Python package. Summary: bhat2023a.
On-the-fly fine-tuning during relaxations outside the training domain (2024)
Wang and coworkers tested the Finetuna framework, in which a pretrained GemNet potential drives a BFGS relaxation and DFT single points are queried at the start, every 30 ML steps and when ML forces appear converged, on systems outside the OC20 domain (Wang et al. 2024). They added a more conservative rule that keeps querying DFT on successive steps until the retrained force error falls below 0.05 eV/Å. With these rules, runs on C3 to C6 alcohol intermediates and on zeolite and MOF structures used under about 20% of the DFT calls of plain VASP BFGS, while spin-polarized oxides needed 36 to 58% depending on the pretrained model. Runs whose ML forces exceeded 5 eV/Å rarely converged, giving a practical stopping criterion. Summary: wang2024.
Bayesian optimization and reinforcement learning in catalysis SDLs (2025)
Orouji et al. contrast Bayesian optimization, suited to discrete, expensive, independent experiments such as choosing catalyst and ligand combinations or temperature, with reinforcement learning, suited to sequential control where the history of conditions matters, such as adjusting gas feed ratios during CO2 hydrogenation (Orouji et al. 2025). They describe hybrids (BO tuning RL hyperparameters) and safe variants that respect reactor limits, and review active-learning campaigns such as a higher alcohol synthesis study that found an Fe65Co19Cu5Zr11 catalyst in about 100 manual experiments from a space of about 5 billion possibilities. Summary: orouji2025.
Sequential versus batch active learning for transition states (2026)
In an active-learning loop of MLP-driven saddle search, one DFT single point on the result, and force-only fine-tuning of eSEN-OAM, a sequential strategy that specializes a model to each transition state needed on average about 8 DFT single points per structure, compared with about 38 for a batch strategy that shares one model across all searches, about 70 for DFT optimization after MLP pre-optimization and about 102 for full DFT optimization (Cheula et al. 2026). The sequential loop matched reference NEB transition states (within 0.1 eV) as often as full DFT; the authors note its efficiency comes from deliberate overfitting to a local region of the energy surface, while the batch model is more transferable. Summary: cheula2026.