Expert background

• 5+ years at BASF; Responsible for enabling the digital platform for materials scientists and facilitating their research work
• Not a scientist; focused on the digital platform

Current &D Process

• First, start with a new property you are searching for. Then with an LLM, take training data to put into the model to output 800-1,000 possibilities
  • Out of the 1000 possibilities, you need to use a platform to trim them down due to safety concerns and stability concerns; trim down to 100
  • We have pretty high precision because we have a long history of experiments and have that all in a database
  • These 100 possibilities are then passed on to the scientists for further investigation
• Timelines
  • New molecule computational enhancement used to take 6 months - 1 year, but with the new Curiosity platform, seeing a 15-day to 1-month scenario
    • Typical scenario: with 40k possibilities, our new platform can parse through it in a minute, which used to take a year
  • Multi-scale modeling + engineering modeling takes 2.5 - 3 months for a typical additive for new model

BASF strategy on deciding where to focus R&D

• For enhancements: We capture a lot of NPS data and decide if we need an enhancement because we’re now losing to a competitor. It’s not based on a fixed schedule but rather on customer feedback
• We do a fixed schedule for new molecules
  • E.g., for battery solvent: invention is more and more frequent because there is a lot of appetite from China
  • But for catalysts, the new search is much less frequent (sunsetting it) because it will go away over the next 10 years

BASF Curiosity Platform

• 1.75 petaflop compute, 200 tb RAM, 5 petabyte storage. HPC platform
• Applications
  • Some are in-house, we also have an R&D app store
• AI Modelw
  • Crystal AI https://crystals.ai/
  • ML Flow
  • BASF has not invested into LLMs to-date. We are integrating Open AI models with our data, but there are IP restrictions. We haven’t developed our own models but will in the future

Perspective on competitive threat from big technology platforms

• Google/Meta in the long run won’t be able to do niche, but will provide the infra for this
• Will have players who build niche-specific models for chemistry
  • Crystals AI is one
  • For DF modeling there are many vendors, including some who are AI based
  • In BASF, we have 45k products. For use cases regarding finding new molecules + enhancements, it’s super fragmented. Then also you have oil and gas, and other segments. Also different types of molecules. BASF users will need to fine tune models for each part of this matrix