Quanscient.allsolve: Reducing Multiphysics Simulation Times from Days to Coffee Breaks

There’s a joke among computational scientists that as long as you can run a simulation overnight and get the result the next morning, you’re fine. In reality, though, it almost always takes longer, oftentimes significantly longer! 

Whether you’re developing a fusion reactor or a novel type of microchip, nowadays, you start developing a complex hardware product by building a simulation model first. This prevents wasting a lot of resources on prototyping. However, these models quickly become complex and hard to run, eventually taking days, weeks, or even months to produce results. 

Quanscient is making waves in the industry with its multiphysics simulation toolkit that harnesses the power of distributed computing and quantum computing to shorten the duration of simulations by more than 100 times, shaving months off the product-to-market time. After our first interview in the summer of 2022, they closed a €3.9M financing round led by Maki.vc and got their first product, Quanscient.allsolve, to the market.

In this expert interview, we discuss product roadmap, upcoming milestones, and fundraising for deep tech startups with the CEO, Juha Riippi, and Chief Scientist, Valtteri Lahtinen

What’s Your Product Strategy to Address Various Verticals?

When we spoke one year ago, we were pre-product but already had a few pilot projects with potential customers ongoing. Since then, we launched our first product, Quanscient.allsolve, focusing on two verticals: superconducting technologies and micro-electromechanical systems (MEMS). MEMS are tiny integrated chips that combine mechanical and electronic components. 

Our simulation software can, in principle, already do many things, like designing antennas, semiconductor chips, quantum devices, electric motors, or even superconducting electric motors, with applications in aerospace, for example. 

For MEMS and superconductors, we already knew the industries. Valtteri did his Ph.D. on modeling superconductors, while our CTO, Alexandre Halbach, worked on MEMS design at imec, a Belgium semiconductor fab. 

Superconductors are used for developing so-called superconducting qubits, one of the most promising hardware platforms for building quantum computers, and a well-established use case that allowed us to gather customer feedback early on. They also come in handy for building fusion reactors, where the challenge is to design superconducting coils so that they generate a magnetic field that confines the fusion plasma in a desired way—this is very non-linear by nature and extremely difficult to do with existing tools. In collaboration with researchers from MIT, we’ve demonstrated one simulation for a superconductor, where Quanscient.allsolve brought the simulation time from eight days to three hours. This is detailed in the original paper and on our Quanscient’s blog). Both use cases are really R&D heavy but equally inspiring: our simulation tools could contribute to making quantum computing and fusion energy a reality one day. How cool is that!

The MEMS industry is more established, and it’s less about R&D and more about optimizing devices for production—both to avoid device failure and ensure nothing goes wrong during mass production, as well as optimizing the device design. So far, people have used simulations mainly in the initial design phase of a MEMS device but not close to production. At that point, they don’t want to make major changes. But having access to fast simulations makes it economical to run hundreds of thousands of small simulations, to test little variations in the design and find an optimal one. Even just 1% savings from design optimization can be a big deal at the scale of semiconductor manufacturing. 

How Will You Develop Quanscient.allsolve Further?

We’re employing a software-as-a-service business model, so our pricing is not constrained by licensing and depends only on usage. This allows our users to use simulations flexibly, for instance, running small simulations very fast and many of them or running very complex simulations at scale, both accelerated by distributed computing.

We’re running our simulations on traditional CPUs on AWS since our classical algorithms are based on finite-element methods (FEM) that are very memory-intensive. GPUs typically have lots of cores but limited memory, and while some FEM algorithms can benefit from GPUs, we can already achieve tremendous performance with CPUs—and they are also cheaper than high-end GPUs. FEM simulations are also still too niche to design application-specific integrated circuits (ASICs) tailored to them. 

With our recent seed funding, we want to validate product-market fit for superconductors and MEMS and build the best and most capable simulation platform. We’re already talking to about 30 clients, including some big ones, and a million in annual recurring revenue (ARR). In 12-18 months, we will also release an API to allow, for instance, large MEMS companies to integrate our solvers into their workflows without using our graphical user interface. 

As we’re in a unique position to gather lots of simulation data over time, we could leverage them to train machine learning models. Using the predictions of these models, we can create shortcuts for simulations, reducing simulation times further by order of magnitudes. It could also help to create an accurate digital twin, for example, of an electric motor, and evaluate real-world sensor data for its maintenance. Finally, machine learning can also give feedback to engineers to help them optimize their designs in the first place, for example, optimize MEMS devices for material consumption.

In addition, over the next two years, we are going to leverage quantum computing to get a quantum advantage, at least for very specific use cases, such as in computational fluid dynamics (CFD). It’s a critical use case since up to 70% of supercomputing time globally is spent on, for example, aerodynamics or climate simulations with CFD at their core. We developed our own patents and quantum algorithms for CFD, implementing the quantum lattice Boltzmann method, which is purely a quantum algorithm. There is a live interactive demo which people can sign up for. 

We used this quantum algorithm to run CFD simulations on a real quantum device and demonstrated that as the size of the CFD simulation grows, the number of required operations grows only logarithmically, which indicates an advantage. Of course, the problem sizes that quantum computers can handle are still small, but as they become more powerful, we want to connect our classical CFD solvers and our quantum algorithms. 

How Did You Raise Funding in the Current Climate?

The good thing about being all first-time founders is that we don’t have experience with the easy times in startup life, so we put a lot of systematic effort into building Quanscient. Luckily, the fundraising process with Maki.vc went pretty smoothly for our latest funding round: they’re specialists and grasped from the start what we’re doing.

Deep tech funds typically have people who are familiar with multiphysics simulations, and you can discuss with them on a technical level. With generalist funds, it’s more tricky as you need to explain a lot of the fundamentals. And it’s hard to explain those in an elevator pitch. We talked a lot to investors, but focusing on deep tech investors with awareness of the use cases we’re addressing worked rather well. We will start fundraising again at the end of 2023.

Comments are closed.