The PEDS models require significantly less training data, reducing it by at least a factor of 100 to achieve a target error of 5 percent. This new method could potentially accelerate simulations of complex systems in various fields such as weather forecasts, carbon capture, and nuclear reactors. The findings were detailed in the journal Nature Machine Intelligence.
Key takeaways:
- Researchers have developed a new approach to solving complex equations more efficiently using brain-inspired neural networks, which could have numerous applications in science and engineering.
- The new method, known as physics-enhanced deep surrogate (PEDS) models, uses physics simulators to train neural networks to match the output of high-precision numerical systems.
- PEDS models were found to be up to three times as accurate as other neural networks at tackling partial differential equations, and required only about 1,000 training points, reducing the training data required by at least a factor of 100.
- Potential applications for PEDS models include accelerating simulations of complex systems in engineering such as weather forecasts, carbon capture, and nuclear reactors.