3IA PhD/Postdoc Seminar #53

Flash presentation | Machine Learning for Sample-Efficient Molecular Property Prediction

Abstract: In the last decade, there has been a significant adoption of machine learning (ML) techniques in molecular sciences. In particular, supervised learning methods with strong inductive bias were developed for molecular property prediction. Usually, computing these properties require solving complex systems of equations numerically, which could take weeks to complete. On the other hand, with ML models, the computational time is greatly reduced, provided a comprehensive dataset with pre-computed properties is available. In my PhD, I developed a sample-efficient method based on active learning to train these models using as few data samples as possible. A computationally-efficient software implementation enabled molecular dynamics simulations using the trained models, providing a complete ML pipeline for end-users. My contributions paved the way for further developments in machine learning integrations in molecular sciences.

CNN-Augmented Fluid Solver for Capturing Fine-Scale Turbulence

Abstract: Understanding turbulence is crucial for studying gas-dust dynamics in planet forming disks, yet resolving small-scale structures remains computationally expensive. To address this challenge, we employ a hybrid simulation framework that couples a coarse resolution fluid solver with a convolutional neural network to model unresolved dynamics during time integration. The network operates on solver outputs and is trained in a solver-in-the-loop setting, in which a differentiable fluid solver is integrated directly into optimisation, enabling end-to-end training against downsampled high-resolution direct numerical simulations. To improve long-term predictive performance, the model is trained on extended temporal rollouts. The framework has previously been evaluated on 2D homogeneous isotropic turbulence, where training on extended rollouts significantly improved long-term accuracy. The model also reproduced key turbulence statistics consistent with high resolution reference simulations.
Building on this promising baseline and with the goal to better capture temporal dependencies, we investigate including pre-solver fluid states as additional inputs to the network, effectively encoding information from previous timesteps. We explore two different architectures for integrating this information and compare their performance during inference. Preliminary results indicate that the proposed approach matches baseline performance while requiring substantially shorter temporal rollouts during training, enabling faster optimisation and reduced memory usage.

Solver and Generator of Cryptarithm Using Constraint Programming

Abstract: A cryptarithm is a mathematical and logical puzzle in which words form an equation where the letters represent numbers to be determined in a given base. This problem is popular in recreational mathematics, education, and constraint programming. We propose a general, efficient and easy-to-use approach to solve this NP-Complete problem, as well as a hierarchical approach for their generation. The experimental evaluation has generated a large, diverse, and remarkable collection of new cryptarithms.

Open discussion about all the contributions