Today, Radical AI announces Efficient Geometric Interatomic Potential (EGIP), our first-generation interatomic potential that confirms the performance and speed benefits of direct force prediction.

Historically, physics-based models have driven interatomic potential methods, but the ability to leverage ML fully has been wrongly underexplored. Direct force prediction unlocks unprecedented scalability, helping us to discover revolutionary materials by simulating them at scales that conventional atomistic approaches cannot.

Our EGIP model is only one factor in the full set of ML-techniques that we are using to transform materials discovery and development.

Introducing Efficient Geometric Interatomic Potential

Efficiently and accurately simulating materials at the atomic level is crucial for advancing materials science. Efficient Geometric Interatomic Potential (EGIP) is our first-generation interatomic potential designed for fast and accurate atomistic simulations. EGIP pushes the boundaries of the performance-speed trade-off by employing a combination of strategic architecture design, targeted pruning, direct force predictions, and a refined understanding of precision and graph construction impacts. Our model demonstrates competitive results on key material property predictions, including thermal conductivity, crystal structure stability, and stable crystal structures, paving the way for accelerated material discovery and design.

In contrast to recent model releases that have benchmarked the thermal conductivity task, EGIP proves that direct force prediction models can perform effectively on this task, if built correctly. Not only does our model achieve state-of-the-art results, but it is also the only direct force model ranked in the top five on the Matbench Discovery leaderboard. Direct force prediction models offer advantages over gradient based force models in speed, scalability, and robustness, making them a superior choice for a variety of tasks.

Results on Matbench Discovery

Matbench Discovery is a machine learning benchmark designed to evaluate and rank ML models on tasks that simulate high-throughput discovery of new stable inorganic crystals. Its primary focus is on assessing how well models can identify stable materials, predict their ground-state atomic positions, and estimate thermal conductivity—key challenges in computational materials science. Additionally, we included the inference cost and memory requirements necessary to run simulations on a 100-atom system.

The EGIP model is state-of-the-art on thermal conductivity prediction and is competitive across all metrics while being significantly efficient in computational costs. The table below demonstrates the current ranking of models from various industries (Meta, Orbital Materials, Microsoft) and academic labs. The models are ranked based on a Combined Performance Score (CPS).

Performance vs Speed

For the computation of properties of billions of materials, it is imperative that potentials exhibit speed, scalability, and maintain high accuracy. The subsequent plot depicts inference cost on the x-axis and error in thermal conductivity prediction on the y-axis. The magnitude of memory utilization is represented by the size of the data points. EGIP achieves state-of-the-art accuracy while demonstrating a 3.5x improvement in speed and the ability to simulate systems with 12x more atoms compared to the leading eSEN model.

Model Design and Pruning

Our approach builds on the GemNet architecture, optimized through systematic pruning of components offering marginal performance gains at disproportionate computational costs. Hyperparameter ablation studies revealed three features with minimal impact on model accuracy for this dataset:

• Atomic embedding size (redundant beyond a threshold dimensionality)
• Atom Edge / Edge Atom interaction layers (increases learning capacity with limited accuracy benefits)
• Quadruplet interaction neighbors (computationally prohibitive with limited accuracy benefits)

The refined architecture retains two critical properties:

1. Direct force prediction with strict equivariance under rotation.
2. Physical consistency via hard-coded zero net force constraints (∑F=0).

Notably, quadruplet interactions—which scale as O(N⁴)—were replaced with O(N³) triplet interactions after a targeted fine-tuning phase. The pruning strategy demonstrates that careful component removal, coupled with recovery training, can maintain performance while significantly enhancing inference speed and memory of simulations.

With a combination of efficient model design and pruning we find that our approach improves the inference cost and memory by 52% and 60% respectively while only taking a 4.2% performance hit on the validation metrics of Omat dataset.

Key factors impacting K_SRME metric

We find that direct force models can give competitive results on the thermal conductivity task and the factors that impact the most are neighbor list, displacements, and precision. Note that models trained on fixed neighbors and performing inference at lower inference are much faster even if they are not potentially smooth at smaller displacements.

Machine learning is catapulting industries into a new era, boldly redefining the landscape of science with innovations that once seemed unattainable. EGIP is Radical AI's first public demonstration of a machine learning centric approach. If you are interested in developing the future technologies that power materials science through ML and automation, we encourage you to reach out.