Resumen
Machine learning (ML), particularly artificial neural networks (ANNs) and deep learning (DL), has emerged as a powerful tool for analysing complex experimental data. However, traditional supervised ML methods are constrained by the need for large paired input-label datasets, which are often difficult to obtain in scientific applications. Furthermore, many ANN models suffer from a lack of interpretability, raising concerns about their alignment with physical principles.
This seminar explores the transition from purely data-driven ANN models to hybrid, physics-informed modelling strategies. We discuss how ML techniques are being integrated not only in data pre-processing but also during inference through physicsembedded network architectures. In particular, we highlight methods that incorporate physical models either as priors or through embedding physics directly into neural architectures.
Emphasizing interpretability — particularly in fields like materials science and chemistry — we highlight how physics-informed neural networks provide robust, explainable models that adhere to fundamental scientific laws. By integrating physical insights into ML frameworks, these approaches enhance predictive reliability and improve the understanding of high-dimensional experimental data. This perspective underscores the growing synergy between data-driven learning and physics-based modelling in scientific research.
Ponente
Dr. Ivan Argatov
Biofilms – Research Center for Biointerfaces, Malmö University
Informes
luis.lopez@mym.iimas.unam.mx
daniel.castanon@iimas.unam.mx
Actividad presencial con transmisión a través de Zoom, previa inscripción en https://shorturl.at/jq1Ak