A team of researchers at EPFL’s Intelligent Maintenance and Operations Systems (IMOS) laboratory has developed a groundbreaking AI algorithm called Dynami-CAL GraphNet that ensures simulations of physical systems remain stable and physically accurate by embedding Newton’s third law directly into its architecture. The work, published in Nature Communications, addresses a major challenge in AI: while machine learning models can make impressive predictions, they often fail to respect fundamental physical laws, leading to unrealistic or divergent results over time. Traditional AI models rely on statistical patterns and can produce incoherent outcomes when simulating dynamic systems like colliding particles, robotic movements, or human gait. In contrast, classical physics-based simulations are accurate but computationally expensive and inflexible—requiring complete reconfiguration for new scenarios. Dynami-CAL GraphNet bridges this gap by combining the efficiency of AI with the reliability of physical laws. The algorithm is a graph neural network (GNN), where objects are represented as nodes and their interactions as edges. What sets it apart is that Newton’s third law—every action has an equal and opposite reaction—is built into the model’s core. This ensures that forces between interacting objects are always balanced, preventing the accumulation of errors that plague other AI systems. “Instead of letting the AI guess how physics works, we design it to follow the rules,” says Vinay Sharma, a Ph.D. student involved in the project. “This allows the model to generate stable, credible predictions even in situations it has never seen before.” Because Newton’s third law applies universally across scales and systems, the algorithm can generalize and extrapolate effectively. It maintains stability for over 16,000 consecutive simulation steps—far longer than most AI models—while accurately modeling complex behaviors. The researchers tested Dynami-CAL GraphNet in several real-world scenarios. In one, it simulated thousands of granular particles in a rotating mixer, extrapolating from just four simple training simulations involving dozens of particles in a stationary box. The model successfully predicted the dynamics of large-scale, complex systems with moving boundaries. It also accurately predicted human walking patterns using only basic motion capture data, without being told how much force the ground exerts. At the molecular level, the algorithm successfully modeled protein dynamics in a solvent, capturing subtle deformations over time. A key advantage is that the model requires minimal training data. It can learn to predict future motion from a single time step and then infer entire trajectories. This makes it highly efficient for real-world applications where data is limited. Equally important is the model’s transparency. Unlike many AI systems that operate as black boxes, Dynami-CAL GraphNet computes physical quantities—such as forces, torques, and angular momentum—step by step. These intermediate results are consistent with conservation laws, allowing engineers and scientists to verify the model’s behavior and build trust in its outputs. “Users can check whether the model respects momentum conservation, for example,” says Prof. Olga Fink, head of the IMOS lab. “That interpretability is crucial for safety-critical applications.” By integrating physics into AI design, Dynami-CAL GraphNet offers a powerful new tool for simulating complex systems across engineering, robotics, biomechanics, and materials science—enabling faster, more reliable, and trustworthy predictions.