[{"content":"Martin Uhrin leads the Computational Atomistic Methods and Machine Learning (CAMML) team at SIMaP.\n","externalUrl":null,"permalink":"/members/martin-uhrin/","section":"People","summary":"","title":"Martin Uhrin","type":"members"},{"content":"tensorial is an open-source Python library for machine learning on physical tensors. It combines JAX-based numerical tooling with geometric and equivariant model components, with a focus on atomistic and materials applications.\nIt is built for tensor-valued representations and symmetry-aware learning, which makes it well suited to molecular and materials modelling, graph-based neural networks and other physics-informed machine-learning workflows.\nHighlights:\ntensorial.gcnn: building blocks for equivariant graph neural networks tensorial.mlip: tools for building and training machine-learned interatomic potentials Built on JAX, e3nn-jax and Flax ","externalUrl":null,"permalink":"/codes/tensorial/","section":"Codes","summary":"Machine learning on physical tensors: equivariant graph neural networks and machine-learned interatomic potentials in JAX.","title":"tensorial","type":"codes"},{"content":"Giovanni Trezza is a postdoctoral researcher specialising in atomistic methods.\n","externalUrl":null,"permalink":"/members/giovanni-trezza/","section":"People","summary":"","title":"Giovanni Trezza","type":"members"},{"content":"Mattia Ragni is a PhD student specialising in atomistic simulations and machine learning.\n","externalUrl":null,"permalink":"/members/mattia-ragni/","section":"People","summary":"","title":"Mattia Ragni","type":"members"},{"content":"Zhiyi Wang is a PhD student focusing on machine learning applied to atomistic methods.\n","externalUrl":null,"permalink":"/members/zhiyi-wang/","section":"People","summary":"","title":"Zhiyi Wang","type":"members"},{"content":"Aycha Chouchene is a Master\u0026rsquo;s student in the CAMML team.\n","externalUrl":null,"permalink":"/members/aycha-chouchene/","section":"People","summary":"","title":"Aycha Chouchene","type":"members"},{"content":"Aditi Chauhan is a member of the CAMML team.\n","externalUrl":null,"permalink":"/members/aditi-chauhan/","section":"People","summary":"","title":"Aditi Chauhan","type":"members"},{"content":"Welcome to the Computational Atomistic Methods and Machine Learning team, based at SIMaP in Grenoble.\nWe develop physics-informed machine learning for materials science: equivariant neural networks that respect the symmetries of nature, and generative models that design new materials and molecules with the properties we need. Read more about our research →\n","externalUrl":null,"permalink":"/","section":"CAMML@SIMaP","summary":"","title":"CAMML@SIMaP","type":"page"},{"content":"","externalUrl":null,"permalink":"/categories/","section":"Categories","summary":"","title":"Categories","type":"categories"},{"content":"We release the methods we develop as open-source software, so that others can use, check and build on them.\n","externalUrl":null,"permalink":"/codes/","section":"Codes","summary":"","title":"Codes","type":"codes"},{"content":"","externalUrl":null,"permalink":"/tags/e3nn-jax/","section":"Tags","summary":"","title":"E3nn-Jax","type":"tags"},{"content":"","externalUrl":null,"permalink":"/tags/equivariant-nns/","section":"Tags","summary":"","title":"Equivariant NNs","type":"tags"},{"content":"","externalUrl":null,"permalink":"/tags/jax/","section":"Tags","summary":"","title":"JAX","type":"tags"},{"content":"","externalUrl":null,"permalink":"/tags/mlips/","section":"Tags","summary":"","title":"MLIPs","type":"tags"},{"content":"Meet the people behind CAMML.\n","externalUrl":null,"permalink":"/members/","section":"People","summary":"","title":"People","type":"members"},{"content":"For most of history, new materials have been found the way Edison found his light-bulb filament: by trial and error. Each candidate is made or simulated, tested, and usually discarded, and the space of possible materials is far too vast to explore this way. The CAMML team develops computational methods that change this, combining atomistic simulation with machine learning so that we can predict what atoms will do, and then ask which arrangements of atoms will do what we want.\nPhysics-informed, equivariant machine learning # Standard machine-learning models find patterns in data, but they know nothing of the laws of physics. We build models that do. The behaviour of atoms is governed by fundamental symmetries: a molecule\u0026rsquo;s energy doesn\u0026rsquo;t change when it is rotated or translated, and its forces rotate with it. By building these symmetries directly into the architecture of equivariant neural networks, we obtain models that are faster to train, need far less data, and remain faithful to the physical systems they describe.\nWe are active members of the e3nn community, which develops open-source tools for Euclidean-equivariant neural networks.\nAccelerating quantum-mechanical calculations # Quantum-mechanical methods such as density-functional theory are the workhorse of computational materials science, but their cost limits how many candidate materials can be studied. We use physics-informed machine learning to predict the results of these calculations at a fraction of the cost, from interatomic potentials to electronic properties. A current focus is the search for new spin-crossover materials: compounds that switch between magnetic states in response to temperature, pressure or light, with potential applications in sensors, displays and data storage.\nGenerative models for inverse design # Rather than asking \u0026ldquo;what are the properties of this material?\u0026rdquo;, inverse design asks \u0026ldquo;which material has the properties we need?\u0026rdquo;. We develop generative models that learn the patterns underlying known structures and use them to propose entirely new atomic arrangements, both crystals and molecules, that are statistically likely to have the target properties. This turns materials discovery from a search through candidates into a guided design process.\nMaterials for the green transition # Materials science is a quiet engine of the energy transition. Next-generation metal–air batteries, more efficient catalysts and many other technologies we need to tackle the climate crisis are limited by how quickly we can explore materials space. The methods we develop are aimed squarely at speeding up that exploration.\nWhere we are # CAMML is based at the SIMaP laboratory in Grenoble, in the French Alps, where Martin Uhrin holds a MIAI chair, From Edison\u0026rsquo;s trial-and-error approach to the inverse design of materials and molecules. Grenoble is also home to large-scale experimental facilities such as the ESRF and the ILL, giving us close links between theory, computation and real-world materials.\nInterested in joining us or collaborating? Get in touch with Martin.\n","externalUrl":null,"permalink":"/research/","section":"Research","summary":"","title":"Research","type":"research"},{"content":"","externalUrl":null,"permalink":"/series/","section":"Series","summary":"","title":"Series","type":"series"},{"content":"","externalUrl":null,"permalink":"/tags/","section":"Tags","summary":"","title":"Tags","type":"tags"}]