Multiverse Computing: Shaping the Future of Quantum-Inspired Computing
Often hailed as the next frontier in computing, making quantum computing practically and commercially viable remains challenging.
Still, we can gain substantial value from quantum physics right now by applying its principles to develop better classical, so-called ‘quantum-inspired’ algorithms, offering improved efficiency and performance without the need for actual quantum hardware. Instead, they run on traditional microchips, mostly GPUs or even more specialized chips, to achieve faster computation times and address real-world problems already today.
Multiverse Computing leverages quantum-inspired algorithms to increase the performance of machine learning models and find better solutions for optimization problems. Founded by Enrique Lizaso Olmos, Román Orús, Alfonso Rubio-Manzanares, and Samuel Mugel in 2019, Multiverse Computing raised a €25M Series A funding round this spring 2024, led by Columbus Venture Partners followed by Quantonation Ventures and new investors European Innovation Council Fund, Redstone QAI Quantum Fund, and Indi Partners.
Learn more about the future of quantum-inspired computing from our interview with the founder and CTO, Samuel Mugel:
Why Did You Start Multiverse Computing?
When I started my PhD in 2013, researching ultracold atoms in Spain, quantum computing was still an academic curiosity, and most researchers didn’t expect it to be anything more than sci-fi any time soon. By the time I defended my thesis in 2017, quantum computing was commercially available by D-Wave, and the first results had been demonstrated.
That’s when I decided to get involved with it. My co-founders and I started in a fairly non-standard way, first as a non-profit, publishing academic articles and evaluating the potential of quantum computing. Once we had convinced ourselves how we could unlock its massive potential, we founded Multiverse Computing as a company. We were approached almost immediately by Quantum Stream at Creative Destruction Lab in Toronto, and through the program, we zeroed in on our value-add and captured our first North American customers.
How Does Quantum-Inspired Computing Work?
At the moment, about 25% of our work is on quantum computing itself, while the majority of our work is ‘quantum-inspired.’ It means that we leverage insights and mathematical methods from quantum physics to develop better classical algorithms, mainly for machine learning and optimization. While they currently run on GPUs, we’re fully prepared to take advantage of quantum computers as they become more powerful.
Regarding machine learning, our main goal is to achieve better model performance by leveraging quantum methods like tensor networks. It’s a mathematical method that’s similar to ‘principal component analysis’ in the sense it determines which parts of a system represent most of the information content and which parts are redundant. This insight allows us to compress a model to make it run faster and reduce inference times. Customers can access this as a product through our own cloud or license the technology to deploy it on-premise.
Regarding optimization problems, our mission is to outperform solutions obtained with established tools like Gurobi. We use standard optimization strategies and develop custom, quantum-inspired algorithms in-house. Optimization problems typically come with many constraints that are inequalities rather than equations, and we can help a lot with handling these constraints when solving the optimization problem.
Interestingly, constrained problems are often more challenging to solve than unconstrained ones, though intuitively, it shouldn’t be that way. The constraints reduce the possible solution space, so intuitively, it should be easier to find an optimal solution. In practice, however, dealing with constraints is tough, and it all depends on how you can simplify the optimization problem despite the constraints, and here, we have an advantage.
What’s Your Product Strategy for Balancing Machine Learning With Optimization and Using Quantum Computing?
You may wonder why we’re tackling two seemingly unrelated domains, machine learning and optimization problems, at the same time. However, if you want to train a machine learning model well, essentially, you need to solve an optimization problem. Thus, working on machine learning led us to work on optimization. Also, from a higher-level point of view, you can treat many problems of our larger customers, e.g., in the chemical industry, either as an optimization or machine learning problem.
Customers don’t care how you solve their problems, but they care about getting the best possible solution today. That’s why we develop our own algorithms and make them run on various computing platforms. We’re already working with GPUs and FPGAs and getting started using ASICs.
Specialized and analog chips can already give an edge to classical computing today. In the end, we think quantum computing will always win—it’s just a matter of how long it takes to get there. That’s because there is mathematical proof that quantum computers outperform classical computers in solving some particular yet relevant problems. I expect we will see the first commercially valuable proof of quantum advantage in less than five years.
Quantum computing makes a lot of sense for chemistry. For example, we have classical algorithms to simulate the ground state of a molecule, but they easily break down for quantum quenches—sudden changes in the parameters of a quantum system, e.g., when you swiftly turn on an external magnetic or electric field. Classical algorithms also break down quantum dynamics and excited states of molecules, which is extremely relevant for catalysis and multi-step reactions. Quantum computing can help us understand these better. We have already seen several demonstrations of quantum advantage in simulating quantum quenches, and it’s just a matter of time before quantum computing shows the first value for chemistry.
As for optimization problems, you’d expect that it will take much longer for quantum computing to show value since you’ll need to handle a lot of data and thus need many qubits. But that’s only the case if you map one variable of an optimization problem to one qubit. But unlike a bit, which takes just two values, zero or one, a qubit can take a continuous range of values, and thus you can map continuous variables to a qubit. Or even several continuous variables; it all depends on how much precision you can control and read out your qubits with in practice. In this setup, solving complicated optimization problems in quantum computing doesn’t necessarily require terribly many qubits: as the Hilbert space grows exponentially with the number of qubits, you just need a few qubits (compared to classical bits) to explore vast optimization spaces.
We’re fully prepared to take advantage of quantum computing as it matures, creating lots of IP, having very good relationships established with all major quantum hardware providers, and building internal teams with a lot of know-how already today.
What Advice Would You Give Fellow Deep Tech Founders?
When I worked on my first startup for one year, I didn’t have any salary and never made a single sale. When it crashed, I was depressed and thought I had wasted a lot of time and energy. But as it turned out, even though my first startup never really amounted to anything, I still gained a lot of experience, learnings, and credibility out of it.
If you take big risks, it’s okay if things don’t turn out as you planned—there’s always something valuable to gain from the experience. Just remember to recognize and appreciate the insights or lessons that come from your endeavors.