ARTIFICIAL BOUNDARY. AI can work across disciplines
| Photo Credit:
Khanchit Khirisutchalual
Over a hundred researchers from across the globe gathered in Seville, Spain, this March for the ‘Science Across Boundaries’ symposium to honour Subra Suresh, a scientist who exemplifies the spirit of interdisciplinarity. There, artificial intelligence (AI) was not merely one topic among many — it emerged as the connective thread in nearly every discussion.
Modern AI did not emerge solely from computer science. It drew heavily from psychology, cognitive science, neuroscience, mathematics and statistical physics. This cross-pollination transformed AI from a pattern-recognition tool into a powerful engine for scientific inquiry.
Built intelligence
Large artificial neural networks are universal function approximators, capable of mapping complex relationships between inputs and outputs across a range of physical phenomena. This allows AI to work across disciplines — the same mathematical machinery can model both protein folding and galaxy formation.
In cosmology, for instance, deep-learning models trained on thousands of simulations have bridged the gap between computationally expensive high-resolution simulations and rough analytical approximations. Even when presented with unfamiliar values for parameters such as dark matter density, they generated plausible results, appearing to capture the underlying physics of gravity and relativity. In quantum optics, AI frameworks such as PyTheus are proposing experimental configurations not known to human physicists.
Markus Buehler and his MIT team presented ScienceClaw + Infinite, a generative AI framework for materials science. Researchers post problems, and AI agents conduct simulations, design experiments and refine models. Infinite extracts scaling laws and builds predictive world models.
The framework, as Buehler described it, acts as a “world-shaping machine” capable of creating materials and engineering structures.
George Karniadakis of Brown University reinforced this vision through physics-informed neural networks, which embed conservation laws into the learning process. By incorporating physical constraints, these systems can learn even from sparse or noisy data.
Technological capability without pedagogical wisdom risks producing tools we cannot responsibly wield. Traditional lectures are becoming less effective since information is instantly accessible. So, what unique value do human educators provide?
An experiment at IIT-Madras offered an answer. AI systems analysing why students failed programming examinations found that the issue was not merely syntax errors. They identified multiple categories of misunderstanding, ranging from debugging difficulties to flawed algorithmic logic. This helped in creating personalised tutorials suited to individual needs.
Learning anew
Curricula, too, must adapt. At the IIT-Madras Wadhwani School of Data Science and AI, undergraduate education uses a “data-first” approach that encourages students to tackle problems through computational and analytical thinking rather than traditional academic silos.
Assessments also require rethinking. Instead of banning AI tools, educators may need to integrate them into assignments — for instance, asking students to compare conventional programming with AI-assisted methods.
Some of the most sobering discussions at the Seville symposium concerned AI’s societal effects, such as reinforcement of harmful behaviour. Research suggests that individuals who behave aggressively online may become even less likely to apologise if AI systems validate their hostility.
Safekeeping society
Through the IIT-Madras Centre for Responsible AI, researchers are examining how AI reshapes society.
Regulation alone cannot ensure safety. AI literacy must begin early, perhaps in middle school, enabling students to critically evaluate the capabilities and limitations of AI systems. Parents, too, must understand the technology well enough to guide children responsibly.
The circular truth
The symposium illustrated a deeper truth: Boundary-crossing science created AI, and AI now enables boundary-crossing science. Neural networks help physicists understand biology, machine learning allows materials scientists to speak the language of chemists, and generative models connect engineers with quantum theorists.
The boundaries were always partly artificial. AI is making that reality increasingly actionable.
(With inputs from Anil Ananthaswamy and Christos Athanasiou)
(B Ravindran is Professor and Head of the Wadhwani School of Data Science and AI at IIT-Madras; Krishnan Narayanan is President of itihaasa Research and Digital, and a researcher at CeRAI, IIT-Madras)
Published on May 18, 2026
