International Journal of Engineering and Management Research

1. Introduction

Scientific modelling plays a central role in modern science and engineering by providing mathematical representations of natural phenomena. These models enable researchers to understand physical processes, predict system behaviour, optimize engineering designs, and support decision-making in numerous scientific disciplines. Traditionally, scientific modelling has relied on physical laws expressed through differential equations, conservation principles, numerical simulations, and laboratory experiments. Such approaches have successfully solved problems ranging from planetary motion and fluid flow to electromagnetic fields and quantum mechanics. The rapid growth of computational power has significantly expanded the capability of numerical modelling. Nevertheless, many modern scientific problems remain computationally demanding due to their nonlinear behaviour, high dimensionality, stochastic characteristics, and multi-scale interactions. Complex systems such as climate dynamics, plasma physics, turbulence, and biological processes often require enormous computational resources and extensive experimental data for accurate modelling.

Artificial Intelligence (AI), particularly machine learning and deep learning, has recently emerged as a powerful alternative for analysing large datasets and generating predictive models. AI algorithms excel at identifying hidden relationships within data and can often produce rapid predictions once trained. However, purely data-driven AI models suffer from an important limitation: they do not inherently obey the laws of physics. Consequently, AI-generated predictions may violate conservation of energy, momentum, mass, or fail to satisfy initial and boundary conditions, thereby reducing their scientific credibility.

To overcome these limitations, researchers have developed hybrid approaches that integrate established physical knowledge directly into artificial intelligence models. Rather than treating physics and AI as competing methodologies, this integration combines the strengths of both domains. Physics provides theoretical consistency and interpretability, while AI contributes flexibility, adaptability, and computational efficiency.

This paper examines how the integration of physics with artificial intelligence is reshaping scientific modelling.

The advent of experimental and simulation data in scientific research has made artificial intelligence a powerful tool in many scientific disciplines. Data-driven models have some important limitations for scientific applications. They could predict results that would contradict the laws of conservation of mass, momentum or energy, be unrealistic when applied beyond the domain of training, and demand huge amounts of labelled data for learning.

4. Major Applications of Physics-Informed Artificial Intelligence

Physics-based artificial intelligence (PIAI) has emerged as an important computational approach in many different scientific and engineering fields. Through the use of physical laws in conjunction with machine learning, it can provide precise predictions despite lack of experimental data. The subsequent sections highlight some of the important applications of PIAI.

4.1 Fluid Dynamics

Fluid Dynamics is another of the oldest and most productive fields of application of Physics-Informed AI. Conventional methods of Computational Fluid Dynamics (CFD) use complex numerical calculations to solve the Navier-Stokes equations. These calculations become very costly if turbulent flow or multiphase fluids are considered.The main principle of Physics-Informed Neural Network is to incorporate the physical laws which govern the flow into the Neural Network itself. This allows for fast prediction of velocity fields, pressure fields, vortex formation, and heat transfer with much less computation cost compared to conventional CFD techniques.

Physics-informed AI helps scientists in finding approximations to quantum wave functions, solving many body problems, estimation of quantum states, and design of quantum devices.

4.5 Astrophysics and Space Science

Current telescopes produce vast amounts of data which need intelligent processing. AI techniques enable researchers to detect galaxies, categorize stars, discover exoplanets, measure black holes, and analyse gravitational waves. Together with physics equations, predictions will be more accurate and scientifically valid.

4.6 Biomedical Engineering and Medical Physics

Physics-informed AI plays an increasingly important role in healthcare by combining biological physics with machine learning.

4.7 Digital Twins and Smart Manufacturing

Digital twins are the digital models of physical systems updated continually using information from sensors. Physics-informed AI is used to enhance digital twins by adding engineering principles to the models.

medical imaging, geophysical prospecting, and parameter estimation.

5.5 Better Interpretability than Conventional AI

While deep neural networks are typically considered black-box models, the combination of physics helps increase their interpretability. Thus, it becomes possible to check whether the results comply with existing science principles.

6. Comparison with Traditional Computational Methods

The traditional computational approaches and Physics-Informed Artificial Intelligence vary greatly in respect to modelling approach, computing needs, flexibility and prediction capabilities. The former approach is based on physical laws and employs the direct solution of governing equations via FDM, FEM and FVM discretization methods. Despite the high interpretability of results provided by the traditional approach, it becomes costly in terms of computations when dealing with large scale or high-dimensional problems. Physics-Informed AI relies upon the combination of physical knowledge and machine learning to produce predictive models capable of producing approximations much faster once trained. While the initial training stage is computationally intensive, further predictions turn out to be faster than the consecutive run of traditional numerical methods. Moreover, Physics-Informed AI performs well even with a small amount of experimental data as the learning is guided by physical considerations. Nevertheless, the traditional computational approaches retain their superiority in interpretability and robust generalization for well-structured physical domains. In its turn, the performance of Physics-Informed AI becomes limited when used for extrapolation outside the range of training data. For this reason, the future belongs to the hybrid computational schemes combining traditional solvers with intelligent algorithms.

Physics-Informed Artificial Intelligence must not be considered as an alternative to the traditional computational approaches; instead, it should be regarded as an additional tool expanding the abilities of the classical scientific modelling approach. With the help of the combination of theory and data-based intelligence, future scientific models would be more precise, effective, explainable, and flexible enough to adapt to ever more complex reality. Hence, physics and AI integration would be considered the core of next-generation computational science.

References

[1] Karniadakis, G. E., Kevrekidis, I. G., Lu, L., Perdikaris, P., Wang, S., & Yang, L. (2021). Physics-informed machine learning. Nature Reviews Physics, 3.

[2] Cuomo, S., Di Cola, V. S., Giampaolo, F., Rozza, G., Raissi, M., & Piccialli, F. (2022). Scientific machine learning through physics-informed neural networks. Neural Networks.

[3] Mishra, S., & Molinaro, R. (2021). Estimates on the generalization error of physics-informed neural networks. Journal of Computational Physics.

[4] Cai, S., Wang, Z., & Karniadakis, G. E. (2021). Physics-informed neural networks for turbulence modelling. Physical Review Fluids.

[5] Yang, L., Meng, X., & Karniadakis, G. E. (2021). Bayesian physics-informed neural networks. Journal of Computational Physics.

[6] Karumuri, S., et al. (2022). Scalable PINNs via domain decomposition. Computer Methods in Applied Mechanics and Engineering.

[7] Bastek, J. H., et al. (2023). Physics-informed machine learning for quantum systems. Physical Review Research.

[8] Pfaff, T., et al. (2021). Learning mesh-based simulation with graph networks. ICML Proceedings.

[9] Meng, C., et al. (2025). When physics meets machine learning: A survey of physics-informed machine learning. Machine Learning for Computational Science and Engineering.

[10] Toscano, J. D., et al. (2024). From PINNs to PIKANs: Recent advances in physics-informed machine learning.