Physical AI – Selected Works

Overview

Explores the evolution of AI from purely digital, data-driven abstraction toward physically grounded intelligence aligned with the laws of nature. Advances in physics-informed neural networks (PINNs), Neural Physics, and guided generative models illustrate how AI can not only simulate and predict, but also plan and create within real-world physical constraints—bridging digital representations with physical reality.

Keynotes

• Ong, Y. S., Closing the Gap: From Digital AI to Physically Grounded intelligence (Keynote, 32nd International Conference on Neural Information Processing, Okinawa Institute of Science and Technology (OIST), Japan, ICONIP2025, November 20-24, 2025. Details.
• Ong, Y. S., Physically Grounded AI for Scientific Discovery: From Prediction to Generative Design (Invited Speaker, 34th International Joint Conference on Artificial Intelligence (IJCAI'2025), Montreal, Canada, 16 - 22 August, 2025 and Guangzhou, China, 29 - 31 August, 2025). Details.
• Ong, Y. S., AI in Science. (Keynote, Next Generation of Artificial Intelligence Academic Forum, 23rd Conference on International Exchange of Professionals, May 23-24, 2025, Guangzhou, China).

Special Issues

• Zeng, Y., Ong, Y. S., Gupta, A. and Ooi, C. C., Special Issue on Physics-Informed Evolutionary Computation: Advances and Applications, IEEE Transactions on Evolutionary Computation 2025.

Physics-Informed Neural Networks & Neural Physics (Selected Publications)

Selected publications on physics-informed learning and neural physics, including PINNs, evolutionary tuning of physics-informed models, and neural-operator style approaches.

• 2025 JC Wong, A Gupta, CC Ooi, PH Chiu, J Liu, YS Ong Evolutionary Optimization of Physics-Informed Neural Networks: Evo-PINN Frontiers and Opportunities. IEEE Computational Intelligence Magazine, arXiv preprint arXiv:2501.06572 , 2025. Scholar
  Brief: A survey of evolutionary and neuroevolution strategies for improving physics-informed neural models (architecture design / training) for robust neural physics solving.
• 2025 JC Wong, CC Ooi, A Gupta, PH Chiu, JSZ Low, MH Dao, YS Ong Evolutionary Optimization of Physics-Informed Neural Networks: Advancing Generalizability by the Baldwin Effect. IEEE Transactions on Evolutionary Computation, arXiv preprint arXiv:2312.03243. Scholar
  Brief: Introduces evolutionary meta-learning framework for PINNs that learns inductive biases to enable strong generalization across a family of PDE problems and unseen conditions.
• 2025 Z Tan, L Luo, H Yin, YS Ong, W Cai Crowd Dynamics Demand Adaptivity: Self-Adaptive Physics-Informed Neural Network for Crowd Simulation. Proceedings of the 33rd ACM International Conference on Multimedia, 5913-5921 , 2025. Scholar
  Brief: Proposes a self-adaptive PINN that adjusts physics and data signals during training to yield more accurate and physically plausible crowd simulations under varying conditions.
• 2025 X He, L You, H Tian, B Han, I Tsang, YS Ong Lang-PINN: From Language to Physics-Informed Neural Networks via a Multi-Agent Framework. arXiv preprint arXiv:2510.05158 , 2025. Scholar
  Brief: Introduces Lang-PINN, a multi-agent framework that uses language to guide PINN training and enforce physics constraints, improving adaptability and interpretability across tasks.
• 2024 G Jin, JC Wong, A Gupta, S Li, YS Ong Fourier warm start for physics-informed neural networks. Engineering Applications of Artificial Intelligence 132, 107887 , 2024. Scholar
  Brief: Mitigates spectral bias in PINNs via a Fourier warm start that accelerates frequency-wise convergence, enabling smoother optimization and improved accuracy on multi-frequency PDE problems.
• 2023 Z Wei, JC Wong, NWY Sung, A Gupta, CC Ooi, PH Chiu, MH Dao, YS Ong How to select physics-informed neural networks in the absence of ground truth: a pareto front-based strategy. 1st Workshop on the Synergy of Scientific and Machine Learning Modeling@ ICML2023, 2023. Scholar
  Brief: Proposes a Pareto front-based model selection strategy for PINNs that identifies better performing models without ground truth by rescaling losses based on convex Pareto solutions.
• 2023 N Sung, JC Wong, CC Ooi, A Gupta, PH Chiu, YS Ong Neuroevolution of physics-informed neural nets: Benchmark problems and comparative results. Proceedings of the Companion Conference on Genetic and Evolutionary Computation, 2023. Scholar
  Brief: Provides benchmark datasets and open-source code for evaluating neuroevolution methods in physics-informed neural networks, with comparative performance results across standard test problems.
• 2023 JC Wong, PH Chiu, C Ooi, MH Dao, YS Ong LSA-PINN: Linear boundary connectivity loss for solving pdes on complex geometry. 2023 International Joint Conference on Neural Networks (IJCNN), 1-10 , 2023. Scholar
  Brief: Introduces a linear boundary connectivity loss for PINNs that enforces local structure at complex boundaries, enabling accurate PDE solutions with sparser samples and faster training across irregular geometries.
• 2022 PH Chiu, JC Wong, C Ooi, MH Dao, YS Ong CAN-PINN: A fast physics-informed neural network based on coupled-automatic–numerical differentiation method. Computer Methods in Applied Mechanics and Engineering 395, 114909 , 2022. Scholar
  Brief: Introduces CAN-PINN, a fast physics-informed neural network using coupled automatic–numerical differentiation for efficient and physically consistent PDE solving.
• 2022 JC Wong, CC Ooi, A Gupta, YS Ong Learning in sinusoidal spaces with physics-informed neural networks. IEEE Transactions on Artificial Intelligence 5 (3), 985-1000 , 2022. Scholar
  Brief: Shows that mapping inputs to sinusoidal spaces in PINNs increases gradient variability, helping avoid local minima and improving training accuracy across forward and inverse PDE problems.
• 2021 JC Wong, A Gupta, YS Ong Can transfer neuroevolution tractably solve your differential equations?. IEEE Computational Intelligence Magazine 16 (2), 14-30 , 2021. Scholar
  Brief: Proposes a transfer neuroevolution algorithm for solving differential equations with PINNs that leverages prior experience to improve convergence and accuracy across a variety of physics problems.