Beyond the black box

Artificial intelligence is now woven through the fabric of modern science. But scratch the glossy surface and an awkward truth emerges: in many cases, we simply don’t know how these systems reach their conclusions. This closed-nature, often referred to as “black-box”, isn’t a mere philosophical itch, it cuts to the core of how science works, especially in healthcare applications.

Machine-learning models can juggle more variables and data than any human mind could hope to track and process, which makes them enormously powerful. Yet, they may fall short on the core qualities that have underpinned scientific progress for centuries: reproducibility and mechanistic insight. When an AI model offers an answer, we can rarely peek inside to understand the logic behind it. 

Layered on top is a subtler complication: the human fingerprint. Every model is shaped by its creators, by the data they choose, the assumptions they encode and the worldviews they bring. Bias isn’t a glitch, but an inheritance.

So, how do we fix this? In the Journal of Chemical Information and Modelling, VPH member Peter Coveney and Roger Highfield outline a suite of approaches. Explainable AI aims to break open the black box. Causal AI seeks to weed out spurious correlations. But the most promising direction may be hybrid systems in which machine learning is bound together with physics-based modelling.

AI based methods, when combined with physics based models, offer speed without sacrificing rigor, as they are anchored to underlying mechanisms, defined by foundational laws of nature. 

This hybrid vision amounts to a “paradigm shift” in computational physiology, according to Coveney, Highfield, Eric Stahlberg and another VPH member, Mariano Vazquez, in their 2025 Nature Digital Medicine paper.

Mechanistic models like physics-based models already form the backbone of digital twins: virtual replicas of human bodies, or parts of them, built on mathematical equations describing fluid flows, chemical kinetics, and more. These knowledge-based models necessarily simplify the chaotic intricacy of physiology, but they have a critical advantage. Their predictions are transparent, testable and equipped with clear uncertainty estimates. And because they can be personalised using patient-specific data, they move beyond population averages to capture individual biology. On the other hand, mechanisms-based models may at times require substantial computational power to make them work.

AI models like machine learning, meanwhile, excel at detecting intricate patterns and correlations across enormous datasets. AI is proficient at interpolating within the bounds of training data, yet notoriously fragile when pushed beyond those limits. And robust ways of quantifying its uncertainty are critical. In other words, data-driven models deliver answers, often with limited explanations – a critical need for healthcare.

Given the above scenario, the physics-AI partnership can solve the shortcomings of both type of modelling approaches, while magnifying their strengths. Such hybrid strategies are already transforming other scientific areas, from climate modelling and weather prediction to quantum chemistry and simulation of turbulence. Everywhere the pattern is the same: models are faster, more interpretable and grounded in natural laws.

Within healthcare, a remarkable innovation arrived in 2024, when Paula Domínguez-Gómez, Mariano Vazquez, and colleagues published a Nature Digital Medicine study on predicting drug-induced heart rhythm disturbances. The challenge was grand. Evaluating a drug’s interference to the heart’s rhythm or function through virtual clinical trials requires thousands of high-fidelity 3D simulations, a task that requires computing power.

Their solution was elegantly pragmatic. The team first ran 900 detailed physics-based heart simulations, separately modelling male and female anatomy and physiology. Such computational effort required 2.1 million CPU hours. These simulations then served as training data for two machine learning “emulators”, one per sex, offering the potential to predict the risk of arrhythmia at lightning speed.

A single physics-based simulation took 4.3 hours, the emulator needed only a few hundreds of a second, a leap of five orders of magnitude. Running the simulations required nine days to assess a single drug, the emulator delivered the same result in under a second with a 4% average error compared to the simulation.

This simulator-plus-emulator strategy has profound implications. Run the expensive model once, train the simulator, and thereafter generate many predictions at negligible cost, financially, computationally, and environmentally. For drug discovery and early-stage testing, it offers a rapid, accurate screening tool that sharply reduces the barriers to innovation.

And this is only the opening act. As physics-based modelling and data-driven AI continue to converge, they promise a new scientific toolkit that blends speed with understanding, prediction with explanation. The union of these approaches may redefine not just how we simulate the human body, but also how we understand it.

Resources & References:

• Digital twins and Big AI: the future of truly individualised healthcare- npj Digital Medicine, 2025
• Artificial Intelligence Must Be Made More Scientific- Journal of Chemical Information and Modeling, 2024
• Editorial: Integrating machine learning with physics-based modeling of physiological systems- Frontiers in Physiology, 2025 
• Fast and accurate prediction of drug induced proarrhythmic risk with sex specific cardiac emulators- npj Digital Medicine, 2024