# Auto-PINN: Understanding and Optimizing Physics-Informed Neural Architecture

**Authors:** Yicheng Wang, Xiaotian Han, Chia-Yuan Chang, Daochen Zha, Ulisses Braga-Neto, Xia Hu  
**arXiv:** [2205.13748](https://arxiv.org/abs/2205.13748)  
**Date:** 2022-05-27  
**Summary page:** https://www.summarizepaper.com/en/paper/2205.13748v2/

## Key points

- Auto-PINN is a groundbreaking approach for optimizing hyperparameters of Physics-informed neural networks (PINNs) through automated processes.
- PINNs leverage deep learning to solve meshless partial differential equations in irregular and high-dimensional physical domains, revolutionizing scientific and engineering practices.
- Auto-PINN introduces a systematic and automated hyperparameter optimization strategy by incorporating Neural Architecture Search (NAS) techniques into the design process.
- By utilizing NAS, Auto-PINN streamlines the optimization process significantly, eliminating manual or exhaustive exploration of the hyperparameter space associated with PINNs.
- Pre-experiments using standard PDE benchmarks showed that different hyperparameters within PINNs can be decoupled, allowing for fine-tuning specific aspects independently.
- Optimizing based on the training loss function of PINNs serves as a reliable search objective for achieving optimal performance.
- Comparative experiments demonstrated that Auto-PINN consistently outperforms alternative approaches by producing neural architectures characterized by superior stability and accuracy.

## Summary

, , , , 


Auto-PINN is a groundbreaking approach that aims to optimize the hyperparameters of Physics-informed neural networks (PINNs) through automated processes. PINNs have been instrumental in revolutionizing scientific and engineering practices by leveraging deep learning for computational tasks, particularly in solving meshless partial differential equations (PDEs) in irregular and high-dimensional physical domains. The efficiency and accuracy of PINN solvers heavily rely on the neural architecture hyperparameters, presenting a significant challenge due to the vast search space and the complexity of defining suitable search objectives for PDEs. In response to this challenge, Auto-PINN introduces a systematic and automated hyperparameter optimization strategy for PINNs by incorporating Neural Architecture Search (NAS) techniques into the design process. By utilizing NAS, Auto-PINN eliminates the need for manual or exhaustive exploration of the hyperparameter space associated with PINNs, streamlining the optimization process significantly. To validate the effectiveness of Auto-PINN, a series of pre-experiments were conducted using standard PDE benchmarks to investigate the relationship between neural architecture structure and performance in PINNs. The results revealed that different hyperparameters within PINNs can be decoupled, highlighting the potential for fine-tuning specific aspects independently. Additionally, it was found that optimizing based on the training loss function of PINNs serves as a reliable search objective for achieving optimal performance. Comparative experiments with baseline methods demonstrated that Auto-PINN consistently outperforms alternative approaches by producing neural architectures characterized by superior stability and accuracy. This enhanced performance underscores the significance of automated hyperparameter optimization in enhancing capabilities of PINNs for scientific computation tasks. Overall, Auto-PINN represents a significant advancement in optimizing Physics-informed neural network architectures, offering promising prospects for further advancements in scientific computing and engineering applications.