New Open-Source AI Model for Advanced Material Design Unveiled

Sustainable material development specialists Orbital Materials has open-sourced a state-of-the-art AI model for simulating advanced materials to support global decarbonisation.  

Called Orb the model is a fine-tuned version of the company’s foundation model “LINUS” that it has trained from scratch. 

Its launch is part of Orbital’s overarching ambition to leverage AI to help scientists create advanced materials to power the energy transition.

A way to do this is to use computers to peek into the inner functioning of materials and to simulate this activity at an atomic level, something that is impossible to do under a microscope, according to the company. 

“By understanding the mechanisms that give advanced materials their extraordinary properties, we can use computers to design more performant materials,” said Orbital’s CEO Jonathan Godwin.  

But he added that, unsurprisingly, simulating quantum physics is hard. 

“Traditional approaches have been limited by the huge computational resources required, with the consequence being that you can only simulate a drastic simplification of what is really going on at an atomic level,” he said. 

However, he said Orb is more accurate than comparable models from Google and Microsoft and five times faster for large-scale simulations than the leading available alternative.

Related:Meta Unveils Largest Open-Source AI Model In History

Orbital is releasing the model on a permissive open-source license so it is free for non-commercial uses and startups, which Godwin said he hopes will maximize the impact of the technology and accelerate development efforts of teams around the globe. 

He pointed to Orbital as a reminder that a “scrappy, highly motivated startup” can still compete even in an age where AI is so resource, power and people-intensive.

A supporting technical blog is available and a full technical report is planned for the near future to support the use of the model. And the technical description can also be found on the github repository.