GHOST Product Ideas and Venture Scout Conversation Reflections

A 9pm meeting with someone solely focused on money printing got the wheels turning about potential next build avenues as we work towards a room temperature superconductor.

Bryan (the scout) was actually super cool and working out of Hong Kong... this will be more important in a moment.

The way I described GHOST was simply 'ML approaches for high throughput material discovery.' This seemed to immediately resonate with Bryan, and in a way he seemed relieved that I wasn't peddling some new 'AI' modified SaaS. I was candid in framing the work as purely exploratory at the moment, so much of the early conversation centered around the dream.
I forgot exactly how it was articulated but it was something to the tune of: bring transformative tooling to folks that can 'hear the music' (ripping off the Oppenheimer movie), and constantly pursue the discovery of a room temperature superconductor.

We spoke at length about potential business models where I kept coming back to the idea of R&D as a service; taking property requirements and turning around a sample of the material and how to make it. While not the subscription based model he was likely hoping for, Bryan did acknowledge the potential in providing existing industry with new materials prototyping capabilities at fractions of the cost required to pursue such experimentation internally. Bringing back in the Hong Kong bit, Bryan has 'extensive connections' in the Taiwanese manufacturing space. There's certainly no better place on Earth for advanced manufacturing than Taiwan / China at the moment.

Now he kept coming back to subscriptions. I really hated the idea at first, but there's something here (maybe).

Ramblings on a commercialized materials discovery effort:

Making it an experience... imagine a landing page that is a stellar style rendering of constructed latent spaces from recent experimentation (almost like you're flying through space). Minimal boundaries denoting material classes or regions of like-properties. Every point in this space is a valid selection, empty space selections sample the latent space and enable the user to fire up a suite of MLIP / physics based simulations. Simple filters allow researchers to find materials and evaluate their behavior on-demand. No more writing or maintaining in house simulation scripts. Discovery in this lens can be crowdsourced, and the dataset produced by such an app continues to build on the current bleeding edge.

Maybe the individual can play around with it for free. Institutions and industry get fancier stuff for a price.

Then our efforts can focus on providing new tools / research to the user.

Currently we have:

1. VAE or similar latent space fitting experiments

2. A growing understanding of AIMD and the MLIP state of the art

3. Working experience with building predictive tooling on top state of the art MLIP (shoutout @mmoderwell and the ORB Tc predictor)

Currently we need to answer:

1. Do we even care about commercializing (my vote is yes)

2. What is our first 'deliverable' regardless of commercialization (my vote is outline a pipeline that defines latent space sampling -> MD simulations-> physics validations -> property predictors)

3. What should be our development focus (my vote is divide effort between property predictors and latent space construction / sampling)

The conversation with Bryan stoked the fire. Seeing a real path to some transformative work and discovery and I couldn't be more excited.