Expert background

• 20+ years of experience, always in product development
• Started in aerospace on engines and munitions, then went through various BD positions, as well the Apple account. Then led strategy for a $5bn business group. Last six years in digital & AI

R&D process

• Should be similar between us, Dow, Dupont, BASF
• First — Hypothesis generation. I think I can do XYZ. They may have a single or a couple of experiments. Think I can make an adhesive that does this or that
  • Usually, there is little formal structure here. Doesn’t matter where the idea comes from (market or thought on what we could do). E.g., Microsoft came to us looking for a way to make screen feel like paper
  • Not a lot of screening. Screening is not the hard part today, the hard part is following this. This usually just takes 2 or 3 parameters (e.g., need adhesive, feel like paper)
• Second — proof of product.
  • Massive difference from the first stage. Now, I have some chemistry that demonstrates all of the customer requirements. Have a hypothetical understanding that it can be scaled at an economic cost
  • Takes a huge amount of iteration. Need to get a bunch of 10-20 requirements correct (anit-glare, can’t interfere with electronics, durable for multiple years). Success rate to getting a viable product concept is maybe 10% and takes at least 2 years, more realistically 5-10 years
  • It’s not that you have thousands of candidates to screen through. Instead, you need to find an adhesive that can survive for three years. You introduce new parameters and then have to reoptimize. It’s like navigating a 20-dimensional space. In theory, machines have no problem doing this, but experts are very resistant to using new technology
• Third — validated scale and commercialization
  • Doesn’t take that long, 1-1.5 years to build the infrastructure and it’s expensive

Incremental vs. novel

• In incremental, you would have already explored the space and understand more about responsivity. So modestly tweaking the product is faster / shorter extrapolation
  • These are not as interesting from a commercial standpoint because you can’t charge more + cannibalize your own sales
  • If it’s a big performance gain, it is basically the same thing as finding a novel molecule

AI use cases

• The highest-value space isn’t the easiest to go after. This would be a model that is forward predictive enough to say if you take a combo of chemistry and this is what will happen (or even better, do it in reverse, suggest candidates given a set of requirements)
  • However, there’s not a ton of data in materials science. If you look at aerospace adhesive, it was probably built with a few hundred data points
    • Everyone’s data is a mess
    • Anything before the year 2000 is gone. Mainly handwritten data
    • Everything else is in Excel spreadsheets, very esoteric to the person who built it: data schema, logic structure. We were trying to introduce a uniform data platform. We tried to make it look like Excel as much as possible
  • This is where automated labs come in but they’re very specific and expensive
  • There are also significant cultural issues; need a data team of data engineers + software people. These people don’t really work in materials companies, and it’s hard to hire them and retain them because they don’t fit in culturally, etc.
  • Also, it is hard to pay them enough. These people also don’t get to file patents, which is a big requirement for getting promoted in materials companies. Also true at the leadership level
• I don’t think we would be open to a partnership format; our competitive edge is our IP and trade secrets of the production process. Material sciences is our secret sauce. We would think of it as giving away our IP
  • There are some narrow use cases where we’ve done this; we worked with an external party who helped us synthesize, but we ultimately owned the IP. Knew we needed to replace a specific parameter and it was an established business
  • Exclusive royalty partnership model is more palatable; 3M is a bit unique; we are driven by fractional margin, so anything that is dilutive to margin is frowned upon, but this might be specific to us
    • Makes IP portfolio management harder, would box us out from using it elsewhere?
• Intermediate path
  • We spoke to an adhesive developer on how they go about R&D. They had the base formulation and 5 components that would help achieve performance goals. Picked based on gut feeling. Then varied each by 5%, came up with 63 formulations and processed it. Again, based on gut (if it pours like water, I do this. If it pours like maple syrup, I do this). Not based on a model or previous data
  • No system of record for what we’ve done in the past exists yet. Tried an autofill formulation system that shows if others have tried to do what you’ve done already
• What could change?
  • You would need to have a burning need to change something. E.g., A lot of medical adhesives use silicones. They need a release chemistry to be able to peel off the backing of the tape. These were all PFAS but we decided to stop using PFAS by 2025, so we need to reformulate everything
    • Environmental ones don’t seem as urgent
    • Needs to be something that needs to be solved quickly
    • Semiconductors (Antwerp spill lost a bunch of valuable chemicals)
    • See some opps in batteries, innovation there has slowed to be incremental. Carbon fiber composites are not super scalable
    • Pharma already using this type of tooling, very data centric
  • What about MOFs?
    • MOFs have been around and talked about for a long time, but haven’t seen a breakthrough there
    • Feel like the end use cases are quite a while away
      • E.g., don’t see the hydrogen economy becoming a huge thing. Lots of alternatives
    • On carbon, don’t see possibility of an energy benefit, will almost certainly consume energy, which is a cost driver. Could see Shell paying for it because of branding stuff, but unless someone is paying for it, how does it work?

Regulatory approval process