The idea is, you can't really trust off-the-shelf LLMs, predicting the next token, to get complex physical phenomena right. So instead you teach them to use a tool that can model this stuff
Hi all - we’re a new team of ML researchers and chemical engineers trying to predict what the future of chemical process design will look like!
We’re not talking about finding new molecules (i.e. what scientists do) - we’re talking about AI that will help speed up design and scale-up of chemical facilities (i.e. the engineering work). Of course, building things in the real world require SO much caution - but we expect one day designing things in the physical world, without AI supervision, could be seen as driving a car without lane-assist.
We think the important part here is teaching LLMs to use tools, like chemical simulation engines, rather than getting them to predict numbers out of nowhere. This way, any LLM results are verifiable by humans too.
Just as LLMs/VLMs are transforming how we do software engineering and coding, we expect slowly, this trend to propagate into more traditional engineering disciplines.
Would love any feedback! We share a mega-beta v0 in the blog post, too!
