Chemistry and materials research
But there's another problem: simulation is expensive. Expensive enough that, at the moment, when the equation is known and all that stands between us and answers is a big computational slog, industry isn't always willing to pay the costs associated. So they fudge. Run a shortened simulation. Accept an approximate answer.
Background
Industries are aware of limitations in their furnaces, but the lining will wear down eventually. The question is how long it will last. Knowing that earlier is better than knowing it later. Chip makers need to know a year ahead what coatings they'll put on wafers. Nuclear reactor salt eats away at the metal that houses it, and we simulate decades to figure out what will happen.
In many cases, if given the chance, we might simulate twenty different scenarios. Industry simulates one. And then does a real world test, to account for the information that was missed. But by the time they have those results, the opportunity for action has passed and decisions have been made. With corresponding economic losses.
Our four ideas have passed a gauntlet, of sorts. We had fifty five potential ideas, and four made the cut. Some failed because more data are needed, or because it's not always easy to get enough of the right material. Some passed because, even though there are interesting science questions to be answered, the failure mode is compute. Others failed because companies aren't yet investing money on hacks. Lots of interesting science questions were screened out.
Persistent fluorinated compounds are being phased out, and industries need replacements, ideally before deadlines. Ones that work as well, but break down when they're done being useful.
Nobody conducts casual experiments with rocket fuel. Reactor salts take decades to show damage. We simulate decades to figure out what will happen.
Coatings for chips, coatings for ships' hulls, coatings on roofs. Three different industries, all simulated the same way.
All the results include a fingerprint of the setup, which our customer's technical teams can use to validate the results by ensuring the same test setup and reproduce identical results.
Programmes
| Line | The calculation | Who feels it | Stage |
|---|---|---|---|
| Surface coatings | How a coating survives plasma, heat and seawater, at the scale where atoms decide | Chip makers, shipping, buildings | Active |
| Replacing forever chemicals | Degradation and surface behaviour of candidate replacements, on a regulator's clock | Chip makers, manufacturers | Active |
| Dry film lubrication | Solid films under load where no liquid is allowed, in a cleanroom or in vacuum | Fabs, space assembly, robotics | Active |
| Energy and nuclear materials | Molten salt eating metal above 700 degrees, and propellants that cannot be safely tested | Reactor developers, defense | Discovery |
The above test uses one of four Anvelda research lines. While our lines are early stage, the engine underneath them has been tuned in production use.
Benchmark
The above test is a molecular dynamics simulation with 339,695 atoms running on one computer. By tuning the simulation to the target computer chip we increased performance from 4.11 ns/day to 58.90 ns/day. That's a 14.3 times speedup on the same computer, with the same initial structure, simulating the same physics.
An important aspect of each Anvelda benchmark is validation. You might get better performance by cutting physical corners, but that's worthless to us and our customers. In this particular benchmark we ran the simulation with both processors and then compared how far each atom had traveled from its initial position. On average the simulated motion differed by 0.035 angstrom, sitting well inside what thermal fluctuation alone would give you.
That line simulates proteins moving in water, which sounds very different from ionic motion in solid ceramic electrolytes. It's not. At the bottom they are all the same math running on the same computers. We use the same code for our life sciences customers as for chemistry, it's only the customer that's different.
Buyers
Chip making tools experience harsh conditions, including heat and plasma. At those conditions materials fail on an atomic scale, resulting in significant economic losses. Our simulations help companies design coatings and dry films that are more resistant. Those materials are selected a year in advance of fab decisions.
Some materials can only be tested under severe conditions and are tightly regulated, think explosives and propellants, making simulations an economic alternative to physical testing. Defense is a great customer because they buy to a published schedule.
In molten salt reactors the salt is kept above 700 degrees, which corrodes the holding material metal on an atomic scale. Before building one developers need to understand how long the metal will last, which the Department of Energy and defense need simulated decades before it would fail.
There's lots of interest in Anvelda from the industry, but forever chemicals being phased out of chip making also results in purchase orders to fund alternatives, now. Deadlines driven by regulation motivate budgets.
Our company is registered in SAM under UEI: CP9NSKDBXGL7 and has a registry entry at the SBA (ID number is SBC_002681929). This is how we're able to sell to the federal government and we're marked as woman owned and minority owned in SAM which has its perks (i.e. set asides).
Collaborate
I'd like to chat with anyone that runs simulations that are so costly, they shorten them. I work with teams to understand current performance in production and then optimize on their infrastructure and hand over the before/after results, documentation of what I did, a fingerprint of how everything is configured and then they verify all of it themselves. I'm a woman founder in production high performance computing.
My tool will run air gapped because there are use cases in medical, defense, banking, etc. Your team always keeps its code.