Differentiable Physics Engines are re-implementations of classical physics simulators (rigid body dynamics, fluid mechanics, soft body deformation) within automatic differentiation frameworks (JAX, PyTorch, TensorFlow) that allow gradients to flow backward through the entire simulation trajectory — enabling inverse problems ("what initial conditions produced this outcome?"), gradient-based robot control optimization, and end-to-end training of neural networks that include physical simulation as an intermediate computation layer.

What Are Differentiable Physics Engines?

• Definition: A differentiable physics engine implements the same numerical integration algorithms as traditional simulators (Euler, Runge-Kutta, Verlet) but within a computational graph that supports reverse-mode automatic differentiation. This means the gradient of any output (final object position, energy, collision force) with respect to any input (initial velocity, control signal, material property) can be computed automatically.
• Classical vs. Differentiable: Traditional physics engines (Bullet, MuJoCo, PhysX) are optimized for fast forward simulation but treat the simulation as a black box — you can observe what happens but cannot compute how the output would change if you adjusted the input. Differentiable engines sacrifice some forward speed to gain the ability to backpropagate through the simulation.
• End-to-End Integration: By making physics differentiable, the simulator becomes a standard differentiable layer that can be inserted between neural network layers. A perception network can feed into a physics simulator, which feeds into a planning network, and gradients flow through the entire pipeline for end-to-end training.

Why Differentiable Physics Engines Matter

• Inverse Problems: "Given that the ball landed at position X, what was the initial velocity?" Traditional approaches require exhaustive search or sampling (Monte Carlo). Differentiable physics computes $partial x_{final} / partial v_{initial}$ directly, enabling gradient descent to find the initial conditions that explain the observed outcome — orders of magnitude faster than search.
• Robot Control Optimization: Differentiable simulation enables gradient-based optimization of robot control policies by backpropagating through the physics of contact, friction, and articulation. Instead of requiring millions of trial-and-error episodes (reinforcement learning), the robot can compute exactly how to adjust its motor commands to achieve the desired trajectory.
• Material Design: Given a target mechanical behavior (specific stiffness, energy absorption, deformation pattern), differentiable simulation enables gradient-based optimization of material properties, microstructure, or geometric design — directly optimizing the physical outcome rather than relying on heuristic search.
• Neural-Physical Hybrid Models: Differentiable physics enables hybrid architectures where known physics (rigid body dynamics, conservation laws) is implemented as differentiable simulation and unknown physics (friction models, material constitutive laws) is learned by neural networks — combining the reliability of known physics with the flexibility of learned components.

Key Differentiable Physics Frameworks

Differentiable Physics Engines are backpropagation-compatible reality — making the laws of physics a transparent, gradient-carrying layer within the neural network optimization loop, enabling machines to reason about physical causality with the same mathematical machinery used to train neural networks.