Ensemble Learning of Machine Learning Force Fields

The predictive fidelity of machine learning force fields (MLFFs) is often limited by the epistemic uncertainty and architectural biases of individual models. Here, we introduce ensemble learning of machine learning force fields (EL-MLFFs), a model-agnostic framework that utilizes a graph neural network (GNN) to integrate predictions from diverse base MLFFs and exploit their complementary error patterns. We propose two fusion architectures: a computationally efficient direct fusion model and a conservative fusion model enforcing energy-force consistency by construction. Evaluated across chemically distinct settings─including methanol adsorption on Cu(100), an in-house elastin-like peptide benchmark, and additional molecular and materials benchmarks─the framework outperforms the evaluated single-model baselines in force MAEs. Across the tested systems, EL-MLFFs also improves several practical simulation-reliability metrics. However, fusing multiple base models can amplify numerical errors and energy drift in finite-precision simulations, even when energy-force consistency is enforced analytically, highlighting numerical stability as a remaining challenge.