Abstract
Numerical simulations are essential for science and engineering but are often computationally prohibitive. This paper introduces PhysicsGen, a benchmark designed to evaluate whether generative models can learn complex physical laws directly from image pairs. By testing architectures like GANs, U-Nets, and Diffusion models on tasks such as sound propagation and motion prediction, we find that while these models achieve massive speedups (up to 20,000x), they struggle with high-order physical relations. Our results highlight a critical "physical literacy" gap in current generative AI.
Introduction
The recent explosion in generative AI has revolutionized image-to-image translation. However, a fundamental question remains: can these models "hallucinate" real physics? Traditional numerical solvers for partial differential equations (PDEs) are slow and expensive. If generative models could treat physics as a visual translation task—mapping an input configuration to a physical output—they could potentially replace traditional solvers in time-sensitive applications.
Methodology and Dataset
The authors established the PhysicsGen Dataset, comprising 300,000 image pairs across three domains:
- Urban Sound Propagation: Mapping city layouts to sound diffraction patterns.
- Lens Distortion: Simulating optical aberrations on facial images using the Brown-Conrady model.
- Bouncing Dynamics: Predicting the trajectory and rotation of a ball on an inclined plane.
A variety of generative architectures were benchmarked, including Pix2Pix (GAN), VAEs, U-Nets, and several Diffusion models (DDPM, Stable Diffusion, DDBM). These were compared against ground-truth physical solvers to assess both computational efficiency and physical accuracy.
Key Insights
The Complexity Barrier
A primary finding is the "Order of Failure": model performance is inversely proportional to the mathematical order of the physical equation. While models handle linear first-order relations (like constant velocity) relatively well, they struggle with second-order terms such as acceleration or complex wave reflections. For instance, in motion prediction, errors in the Y-direction (gravity/impact) were significantly higher than in the X-direction.
Architectural "Personalities"
Different architectures exhibit unique failure modes:
- GANs (Pix2Pix): Produce sharp, physically plausible images but often misplace objects.
- U-Nets and VAEs: Tend to output blurred results in regions of high uncertainty, leading to physically impossible "ghosting."
- Diffusion Models: Can generate multiple conflicting candidates in a single frame due to their probabilistic nature.
Results
The most compelling advantage of generative models is speed. Simulations that take hundreds of seconds using traditional methods can be emulated in milliseconds—a speedup of approximately 20,000x. However, this efficiency comes at the cost of accuracy, particularly in "Non-Line of Sight" (NLoS) areas where complex diffraction occurs. In the sound propagation task, Pix2Pix and DDBM emerged as the most robust performers, though still prone to errors in complex environments.
Conclusion and Future Work
PhysicsGen demonstrates that while generative models are powerful emulators, they lack the "inductive bias" necessary for true physical correctness. They do not inherently respect laws like the conservation of energy. Future research should focus on physics-informed architectures that integrate physical constraints directly into the loss functions, bridging the gap between visual plausibility and physical reality.
Summary updated