Active Learning for Machine Learning Interatomic Potentials

204 concepts169 papers1 proposal

In computational chemistry and materials science, researchers face a fundamental trade-off between scale and accuracy. To simulate the behavior of materials over meaningful timescales, we need to calculate the forces acting on every single atom at every step of a molecular dynamics (MD) simulation.

Traditional ab initio methods like Density Functional Theory (DFT) compute these forces directly from quantum mechanics. They are highly accurate but computationally crushing, limiting simulations to a few hundred atoms for picosecond timescales. Conversely, classical empirical potentials use simplified mathematical functions to approximate atomic interactions. They are fast enough to simulate millions of atoms, but they fail when chemical bonds break or when materials undergo complex phase transitions.

Machine learning interatomic potentials (MLIPs) bridge this gap. By training neural networks or Gaussian processes on quantum mechanical data, MLIPs can predict atomic forces with near-DFT accuracy at a fraction of the computational cost. However, MLIPs suffer from a critical flaw: they are notoriously poor at extrapolation. If an MD simulation enters a region of atomic configuration space that was not covered in the training dataset, the model's predictions become highly erratic, often causing the simulation to unphysically "explode."

To build a robust MLIP, you need a dataset that covers not just stable equilibrium states, but also high-energy transition states, interfaces, and rare defects. Generating this data beforehand is a guessing game. If you sample randomly, you waste massive computational resources on redundant, stable configurations. If you guess manually, human bias will inevitably miss critical transition states.

The solution to this dilemma is uncertainty-driven active learning. Instead of pre-computing a static dataset, we let the machine learning model guide its own data collection in an automated, iterative loop.


The Active Learning Loop

The core intuition behind active learning for MLIPs is simple: only calculate quantum mechanics when the machine learning model admits it is ignorant.

An active learning workflow typically progresses through four recurring phases:

    [ 1. Explore ] ----> Run MD using the current fast MLIP
          ^                               |
          |                               v
    [ 4. Retrain ] <---- Label with high-fidelity DFT <---- [ 3. Filter ]
  1. Explore: A preliminary MLIP is used to run relatively inexpensive MD simulations across various thermodynamic conditions (temperatures, pressures, or compositions).
  2. Monitor: As the simulation progresses, an uncertainty quantification (UQ) metric continuously evaluates the model's confidence in its own force or energy predictions for the current atomic configuration.
  3. Filter & Label: If the uncertainty remains low, the simulation continues safely. If the uncertainty exceeds a predefined threshold, the configuration is flagged as "novel." The simulation is paused, and the exact electronic energy and forces for that specific configuration are calculated using an expensive, high-fidelity quantum solver (like DFT).
  4. Retrain: The newly labeled configuration is added to the training database, the MLIP is retrained, and the cycle repeats.

By targeting only the configurations where the model lacks confidence, this approach maximizes the information density of the dataset. As reviewed by Paraskevi Gkeka et al., this automated exploration mitigates human bias, forcing the model to hunt for rare but physically vital molecular geometries that are easily overlooked in manually curated datasets.


Quantifying Ignorance: The Uncertainty Engine

The entire active learning framework hinges on the ability to accurately measure epistemic uncertainty—the uncertainty arising from a lack of training data. If the UQ metric is too permissive, the simulation will drift into unphysical territory and fail. If it is too conservative, the loop will trigger expensive quantum calculations for configurations it already understands, destroying the computational savings.

Researchers use a few primary strategies to build this uncertainty engine:

Query-by-Committee (Ensembles)

The most common approach is to train an ensemble, or "committee," of multiple neural networks on the same underlying data but with different initial weight initializations or data splits. Because the models are trained on the same data, they will yield nearly identical force predictions in regions of configuration space they have already mapped. However, when they encounter a novel atomic arrangement, their predictions will diverge wildly.

The variance among the committee's predictions serves as a direct proxy for epistemic uncertainty. Ben Nebgen demonstrated that this query-by-committee approach effectively identifies highly informative molecular conformations, allowing models to approach the "gold standard" accuracy of coupled-cluster calculations while minimizing the generation of expensive training data. Similarly, August E. G. Mikkelsen et al. leveraged neural network ensembles to map the complex, shifting energy landscapes of liquid water-platinum interfaces.

Dropout and Evidential Learning

While ensembles are highly effective, running multiple neural networks in parallel adds computational overhead. To circumvent this, alternative UQ mechanisms have emerged. Ellad B. Tadmor et al. developed Dropout Uncertainty Neural Network (DUNN) potentials, which repurpose "dropout"—a standard deep learning regularization technique—at inference time to approximate Bayesian uncertainty from a single model.

Taking single-model UQ a step further, Han Xu et al. introduced evidential interatomic potentials (eIP). By parameterizing a higher-order distribution over the network’s outputs, an eIP can distinguish between data noise (aleatoric uncertainty) and model ignorance (epistemic uncertainty) in a single forward pass, accelerating active learning without ensemble overhead.


Granularity Matters: Atomic vs. Global Uncertainty

When evaluating whether a configuration is uncertain, looking at the total energy error of an entire system can be misleading. In a system of thousands of atoms, a severe structural breakdown might occur in a localized region—such as a single breaking bond or a localized defect—while the rest of the crystal lattice remains perfectly stable. The global energy error might mask this localized failure.

To solve this, modern active learning protocols monitor uncertainty at the atomic level. Edvin Fako developed an automated training protocol that tracks the predictive uncertainty of individual atoms. If the local uncertainty of even a single atom exceeds a threshold, a high-fidelity calculation is triggered. This granular monitoring prevents localized structural instabilities from propagating and causing catastrophic simulation failures.


Pushing into Rare Territories with Enhanced Sampling

Active learning ensures that if an MD simulation visits an unfamiliar state, the model will learn from it. However, standard MD simulations are inherently trapped by thermodynamics: atoms spend the vast majority of their time vibrating in deep, stable potential energy wells. High-energy transition states or rare phase changes occur on timescales far beyond what standard MD can naturally explore, meaning the active learning loop might never even encounter them.

To force the model to learn the boundaries of its phase space, researchers pair active learning with enhanced sampling techniques.

For instance, Ganesh Sivaraman et al. combined active learning with metadynamics—a method that deposits a history-dependent bias potential to push the system out of local energy minima—to explore highly disordered and fragile liquid metal oxides near their melting points.

Similarly, Jiahao Fan et al. developed the RiD-kit framework, which automates a four-phase iterative workflow using reinforced dynamics. By using the variance of a neural network ensemble to dynamically evolve a bias potential, the system intentionally drives the MD trajectory toward high-uncertainty regions, actively hunting for rare transitions and chemical reactions rather than waiting for them to happen spontaneously.

[ Stable State A ] --(Standard MD trapped here)---> [ Energy Barrier ] ---> [ Stable State B ]
                                                            ^
                                 Enhanced Sampling pushes system over

Why It Works: Real-World Impact

By shifting the paradigm from exhaustive, brute-force data collection to strategic, uncertainty-driven sampling, active learning makes previously impossible simulations routine.

The power of this approach is evident across diverse chemical landscapes:

  • Complex Phase Transitions: Manyi Yang et al. successfully mapped the elusive $\lambda$-transition of liquid sulfur, where it transforms from simple rings into long polymeric chains, an achievement requiring quantum-level accuracy across vast structural changes. Manyi Yang also utilized active learning pipelines to confirm the existence of a liquid-liquid critical point in phosphorus under extreme conditions.
  • Heterogeneous Interfaces: Simulating how water behaves at solid boundaries requires capturing subtle electronic interactions. Xianglong Du et al. used active learning to select diverse atomic configurations across temperatures, accurately reproducing the complex vibrational signatures of the alumina-water interface.
  • Solid Solutions and Alloys: For complex mixtures where the combinatorial space of compositions is massive, active learning allows for targeted screening. Marco Eckhoff et al. bridged the gap between theory and experiment for lithium manganese oxide spinels by using active learning to maintain accuracy across wide temperature ranges and structural phases.

Ultimately, uncertainty-guided active learning transforms machine learning potentials from fragile interpolators into robust, self-improving explorers, unlocking the ability to simulate complex chemistry with quantum accuracy at the scale of materials design.

History & citationVersion 2 · 2026-10-04

Current · Permanent link

  1. Version 2 (latest) · 2026-10-04 · 176 papers
  2. Version 1 · 2026-10-04 · 114 papers
Cite this direction

Latest version (a living page):

@misc{lacuna_uncertainty_driven_active_learning_activ_d34d0358,
  title = {uncertainty-driven active learning (active learning for interatomic potentials)},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/uncertainty-driven-active-learning-active-learning-for-interatomic-potentials/rdr_d34d03581d80445997585da64ddf805b},
  note = {Latest version of a living research direction; cite a version for a fixed text. Accessed 2026-11-19},
}

This exact version (version 2, frozen text):

@misc{lacuna_uncertainty_driven_active_learning_activ_d34d0358_v2,
  title = {uncertainty-driven active learning (active learning for interatomic potentials)},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/uncertainty-driven-active-learning-active-learning-for-interatomic-potentials/rdr_d34d03581d80445997585da64ddf805b/v/2},
  note = {Version 2, frozen 2026-10-04. Accessed 2026-11-19},
}
About

These directions overlap with this one through shared papers.

See something wrong? Report a mistake