A new AI-based method advances solid oxide cell modeling by improving physical consistency and computational efficiency

Researchers from the University of Chinese Academy of Sciences and collaborating institutions have developed a novel hierarchical framework that leverages artificial intelligence (AI) to address a long-standing challenge in solid oxide cell (SOC) engineering. The new Physics-Informed Artificial Intelligence (PIAI) method improves both physical consistency and computational efficiency, opening new opportunities for applications such as real-time performance optimization, degradation diagnosis, and the development of robust Digital Twins.

What’s New?

In recent years, AI-driven modeling has become an increasingly popular tool for accelerating the design and operation of high-temperature electrochemical devices like SOCs. However, many existing approaches rely on purely data-driven "black-box" models or computationally prohibitive high-fidelity physics simulations. This makes them difficult to interpret, generalize poorly beyond their training data, and limits their practical use—especially for real-time control and lifecycle management.

To overcome these challenges, Yibin Zhang, together with co-authors, developed a new PIAI framework specifically tailored for SOCs. The framework introduces a three-tiered strategy that systematically integrates physical laws into machine learning, enabling more transparent, efficient, and trustworthy modeling across scales.

How It Works

The proposed framework works by infusing physical knowledge into AI models at three progressive levels of integration:

1. Loss-level Physics Injection: Governing physical equations are enforced as soft constraints during training, guiding the model to learn solutions that respect conservation laws even with sparse data.

2. Architecture-level Fusion: Physical mechanisms (e.g., electrochemical kinetics) are hardwired into the neural network's structure itself, turning physics into a hard, interpretable prior.

3. Integration-level Correction: Interpretable mechanistic models and AI surrogates interact, with physics governing the residual exchange between them, forming the core of a hybrid Digital Twin.

By disentangling how and where physics intervenes, the framework allows researchers to:

• Systematically ensure predictions adhere to fundamental physical principles.

• Achieve reliable results using smaller or partially labeled datasets.

• Gain clearer insight into the model's decision-making process, enhancing trust.

"This hierarchical taxonomy provides a structured pathway for engineers to select appropriate AI interventions based on their specific modeling fidelity and data availability needs," says Chengzhi Guan, a corresponding author of the study.

Validation and Results

The researchers validated the framework through multiple SOC application case studies cited in the review. For instance, a Physics-Informed Neural Network (PINN) surrogate was used to generate performance maps for industrial-scale cells, achieving a voltage error below 0.513% with an inference time of just 0.5 milliseconds—orders of magnitude faster than traditional simulations. In another example, a physics-informed generative adversarial network (GAN) successfully synthesized electrochemically viable electrode microstructures, demonstrating controllable design under physical constraints. The results show that the PIAI approach can outperform conventional methods in balancing accuracy with speed while offering significantly improved interpretability.

Why It Matters

Many AI models in science and engineering suffer from a lack of transparency and require large amounts of data, which can slow adoption in real-world research and industrial settings. For SOCs—a key technology for green hydrogen production and efficient power generation—this gap hinders the development of reliable Digital Twins for predictive operation and maintenance. By addressing both physical consistency and computational efficiency, this work helps build greater trust in AI-assisted decision-making and accelerates progress toward deployable, lifecycle-aware SOC systems.

According to the authors, "PIAI is emerging as a unifying paradigm that bridges credible mechanistic knowledge and efficient learning... it can underpin trustworthy, real-time, lifecycle-aware SOCs engineering."

What’s Next?

The research team identifies key challenges for future work, such as mitigating spectral bias in neural networks, robustly fusing multi-fidelity data, and improving uncertainty quantification. The outlook includes establishing open benchmarks, validating gray-box Digital Twins with model predictive control in demonstrator units, and ultimately realizing microstructure-resolved, adaptive SOC Digital Twins for predictive maintenance and optimization.

Journal Information
This research, titled "Physics-informed artificial intelligence for solid oxide cells: a comprehensive review of frameworks, applications, and prospects for digital twins", was published in AI Mater.
Full article: https://doi.org/10.55092/aimat20250017