Scaling Quantum Accuracy with Machine Learning Interatomic Potentials

361 concepts275 papers1 proposal

For decades, computational materials science has been trapped in a fundamental dilemma: you can have accuracy, or you can have scale, but you cannot have both.

On one side sit ab initio methods like Density Functional Theory (DFT). By approximating the Schrödinger equation, DFT provides high fidelity across diverse chemical environments, phase changes, and defect structures. However, its computational cost scales cubically ($O(N^3)$) with the number of atoms, restricting simulations to a few hundred atoms over picosecond timescales. On the other side sit classical empirical force fields. By treating atoms as balls connected by idealized springs, they achieve linear scaling ($O(N)$), enabling simulations of millions of atoms. Yet, they fail catastrophically when chemical bonds break or when applied to complex, multi-element systems like high-entropy alloys (Liu et al.).

Machine learning interatomic potentials (MLIPs) resolve this dilemma. By treating the potential energy surface (PES) as a high-dimensional function to be interpolated rather than a quantum mechanical equation to be solved from scratch, MLIPs achieve the accuracy of DFT at a fraction of the computational cost (Liu et al.).

The Core Mechanism: The Principle of Nearsightedness

How can a neural network bypass the laws of quantum mechanics and still scale linearly with system size? The answer lies in a fundamental physical principle: the nearsightedness of electronic matter.

In most condensed-matter systems, the quantum mechanical forces acting on a specific atom are overwhelmingly dominated by its immediate neighbors. As you move further away, the influence of distant atoms decays rapidly. MLIPs exploit this by decomposing the total potential energy of a system ($E_{\text{total}}$) into a sum of local atomic energies:

$$E_{\text{total}} = \sum_{i} E_i\left(\mathcal{R}_i\right)$$

where $E_i$ is the energy contribution of atom $i$, and $\mathcal{R}_i$ represents the local chemical environment—the coordinates and identities of all atoms within a fixed spatial cutoff radius ($r_c$) around atom $i$.

Because the cutoff radius is fixed, the number of neighbors inside $\mathcal{R}_i$ does not grow when the overall system size increases. Evaluating the total energy becomes a matter of running $N$ independent local predictions. This local framing transforms a global quantum problem into a local regression problem, unlocking true $O(N)$ linear scaling. Using this architecture, frameworks like the DeePMD-kit have scaled up to simulate over 100 million atoms with ab initio accuracy (Lu et al.).

Representing Symmetries and Equivariance

A neural network cannot simply ingest raw Cartesian coordinates. If you take a molecule and rotate it in space, its total energy remains identical, but its Cartesian coordinates change completely. To learn an efficient mapping, the network must respect the fundamental physical symmetries of space: translation, rotation, and permutation invariance (swapping the labels of two identical atoms should not change the energy).

Early MLIPs solved this by transforming Cartesian coordinates into hand-crafted, invariant geometric descriptors—such as angles and distances—before feeding them into standard regression models (Gubaev et al.). Modern architectures, however, rely on Graph Neural Networks (GNNs) to learn these representations dynamically (Reiser et al.). In an atomistic graph, atoms are nodes and spatial relationships are edges. Open-source toolboxes like SchNetPack leverage message-passing algorithms to propagate structural information across neighboring nodes, building up complex representations of the local environment (Schütt).

The state-of-the-art has advanced further through equivariant neural networks, such as NequIP and EquiformerV2 (Phuthi et al.; Liu et al.). While energy is a scalar (invariant to rotation), forces are vectors (which must rotate exactly as the geometry rotates). Equivariant networks pass directional geometric tensors through their layers without discarding orientation information. This allows the model to predict both energies and directional force vectors with far greater data efficiency.

Exploring the Surface via Active Learning

An MLIP is an interpolator; it is highly accurate inside the domain of its training data but prone to wild errors when extrapolating into unseen atomic configurations. If a molecular dynamics simulation enters an un-sampled region of the potential energy surface, the model may predict non-physical forces, causing atoms to fuse together or fly apart.

To build robust potentials, researchers employ active learning loops, often described as "data flywheels."

 ┌────────────────────────────────────────────────────────┐
 │                                                        │
 ▼                                                        │
┌──────────────────────┐     ┌──────────────────────┐     │
│  Generate Candidate  │────>│  Evaluate with MLIP  │     │
│    Configurations    │     │  (Fast Exploration)  │     │
└──────────────────────┘     └──────────────────────┘     │
                                        │                 │
                                        ▼                 │
                             ┌──────────────────────┐     │
                             │  Uncertainty High?   │     │
                             └──────────────────────┘     │
                                   │          │           │
                               Yes │          │ No        │
                                   ▼          ▼           │
                       ┌──────────────┐   ┌───────────┐   │
                       │ Calculate    │   │ Accept    │   │
                       │ Quantum DFT  │   │ Trajectory│   │
                       └──────────────┘   └───────────┘   │
                               │                          │
                               ▼                          │
                       ┌──────────────┐                   │
                       │ Retrain MLIP │───────────────────┘
                       └──────────────┘

Instead of trying to guess every possible atomic arrangement beforehand, the model is trained on an initial dataset and then set loose to run simulations. Advanced sampling techniques like metadynamics are used to deliberately push the system toward rare events, such as chemical reactions or phase transitions (Yang et al.).

During exploration, the framework monitors the model's prediction uncertainty (often via an ensemble of networks). If the uncertainty stays low, the simulation proceeds rapidly. If the atoms enter a novel configuration and uncertainty spikes, the simulation pauses, a high-fidelity DFT calculation is performed for that specific structure, and the new data is fed back to retrain the potential (Yang et al.). Training models on these full relaxation trajectories yields high zero-shot accuracy across diverse chemical spaces (Merchant).

Unlocking New Frontiers

By combining the speed of empirical force fields with the precision of quantum mechanics, MLIPs have turned previously intractable simulation problems into routine calculations.

Phase Transitions and Extreme Conditions

Simulating structural phase transitions requires modeling thousands of atoms over nanosecond timescales to overcome finite-size effects and capture emergent long-range correlations. Gigli et al. utilized MLIPs to successfully model the ferroelectric phase transition in barium titanate, resolving long-standing discrepancies between simulation and experiment. Under extreme environments, Chen demonstrated that deep potentials can accurately map the complex solid and liquid boundaries of tin across vast pressure and temperature ranges.

Disordered Systems and Ion Transport

In materials like solid-state battery electrolytes, ion transport is heavily influenced by structural disorder. Tracking the diffusion pathways of lithium ions requires long trajectories that are impossible with direct DFT. Zhong et al. and Chen et al. used MLIPs to reveal how local structural and cation disorder creates "soft" local environments that dramatically accelerate ion hopping kinetics.

High-Throughput Screening

In catalyst and superconductor discovery, searching through thousands of candidate crystal structures or molecular adsorption orientations is traditionally bottlenecked by the cost of structural relaxation. Frameworks like AdsorbML use MLIPs as ultra-fast surrogate relaxers to pre-screen and rank thousands of configurations, achieving a 2000-fold speedup while maintaining a near-90% success rate in finding global energy minima (Lan et al.; Dam et al.).

By shifting the computational bottleneck from solving differential equations to evaluating neural networks, MLIPs have fundamentally expanded the scale, duration, and complexity of systems accessible to quantum-accurate molecular simulation.

History & citationVersion 2 · 2026-10-04

Current · Permanent link

  1. Version 2 (latest) · 2026-10-04 · 277 papers
  2. Version 1 · 2026-10-04 · 321 papers
Cite this direction

Latest version (a living page):

@misc{lacuna_machine_learning_interatomic_potentials_7cc2356b,
  title = {machine learning interatomic potentials},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/machine-learning-interatomic-potentials/rdr_7cc2356bf6234b5a953cbc17f7573da4},
  note = {Latest version of a living research direction; cite a version for a fixed text. Accessed 2026-11-20},
}

This exact version (version 2, frozen text):

@misc{lacuna_machine_learning_interatomic_potentials_7cc2356b_v2,
  title = {machine learning interatomic potentials},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/machine-learning-interatomic-potentials/rdr_7cc2356bf6234b5a953cbc17f7573da4/v/2},
  note = {Version 2, frozen 2026-10-04. Accessed 2026-11-20},
}
About

These directions overlap with this one through shared papers.

See something wrong? Report a mistake