The boundary between physics, mathematics and AI is becoming increasingly productive

Logistics can be a logistic nightmare, where small glitches can have a big impact. Take the COVID vaccine, for example. The mRNA-based vaccine, developed in record time during the pandemic, relied on a very strict cold chain. Some doses needed to be kept cold at -70 degrees, for example, and a single temperature change could make a batch unusable. This proved to be quite a challenge: poor route planning, limited storage for the ultra‑cold vaccines, repackaging and sending large multi‑dose vials to clinics that could not use them in time meant that there was both a lot of waste and some dramatic shortages. What of this could be avoided with better planning thanks to an intelligent, robust and adaptable logistics chain?

Particle physicist Nicola Serra from the University of Zurich and mathematician and Fields Medalist Alessio Figalli from ETH Zurich think they might have a solution. Equipped with a 1.7-million-CHF BRIDGE Discovery Grant from the Swiss National Science Foundation and Innosuisse, they have set out to improve the way things like vaccines reach their destinations. The use of a novel combination of a mathematical framework known as “optimal transport theory” and AI methods inspired by particle physics for faster, more efficient and resilient data handling.

So, if the cooling system of a vaccine lorry breaks down or demand rises in an unexpected place, companies could use their system to find or even foresee fast alternative solutions to keep the supply chain from breaking down.

We have spoken to Nicola Serra to find out more about their project.

Question: Your background is in particle physics, meaning you have worked with big data about tiny particles. How did the collaboration with Alessio Figalli start and what motivated you to transfer your physics and maths knowledge to the “real world”? Was there a specific moment or trigger?

Nicola Serra: Alessio and I first started working together in 2022 in a Sinergia project with colleagues from epidemiology and chemistry at the Univerity of Zurich and ETH Zurich. That collaboration made it clear that there is strong value in combining rigorous mathematics with machine learning when working on high-stakes problems. In parallel, interacting with industry, we repeatedly saw that many real-world decisions rely on AI systems that are either opaque or not robust when conditions change. That was a trigger for us. Coming from particle physics, where huge datasets, uncertainty, and real-time decisions are the norm, I realised that many of the tools we use at CERN could be transferred to problems that affect society directly, such as logistics. And supply chains are a good example of something you only notice when disruptions suddenly remind you of how critical they are.

Question: Can you explain in simple language what optimal transport theory is and how it helps finding the most efficient way to move goods and resources worldwide?

Nicola Serra: “Optimal transport” goes back to Gaspard Monge in the 18th century, who was interested in very practical engineering problems: how to move building materials, like piles of earth or bricks, from one place to another with the least possible effort. The modern theory keeps the same spirit but generalises it enormously. It gives a precise way to compare two situations and calculate the minimal “work” needed to turn one into the other.

In our project, we use this idea not to move goods physically, but to measure how much reality is drifting from what a model predicts. That gives the AI a reliable signal for when to adjust its decisions, which makes the system more robust when conditions change.

Question: What is a typical supply chain vulnerability? Can you give a concrete example?

Nicola Serra: One typical vulnerability is that many supply chains are tuned for efficiency rather than resilience. They assume that tomorrow will look like yesterday. As soon as one element shifts - a delayed shipment, a sudden demand spike, a new regulation, or even switching from diesel to electric trucks - the whole system can end up in a fragile state. A concrete example is the shortage of certain pharmaceutical ingredients during the pandemic: a single disrupted supplier affected downstream production, distribution, and ultimately patient access. Similar issues appear in freight transport when a key route is blocked or overloaded, for example the Suez Canal.

Question: What are “simplified simulations” and how does the system learn and adapt in real time to changing supply chain conditions?

Nicola Serra: A simplified simulation is a fast, deliberately approximate digital model of a supply chain or logistics network. Instead of constructing a single ultra-detailed “digital twin”, which is costly and still never fully accurate, we create many simplified simulations at once, each reflecting a different plausible situation. The AI learns by training across this whole family of scenarios, so it develops strategies that remain effective not only in a perfect model but also under a wide range of real-world conditions. We then combine these simulations with real data and the optimal-transport tools described above, which lets the system adjust its decisions as soon as reality starts to shift, without needing to be rebuilt from scratch.

This approach is strongly inspired by particle physics, where we routinely use ensembles of fast simulations to explore possible detector conditions or rare events, and then calibrate them with real data to make our algorithms robust.

Question: How will your approach change the way companies respond to unexpected vulnerabilities? How can they get access to your project?

Nicola Serra: Our aim is to give companies a decision-making tool that remains reliable even when disruptions arise that were never seen in past data. Rather than reacting after a crisis, the system adjusts routing, inventory, or sourcing decisions as soon as it detects that conditions are shifting. Several companies have already expressed interest in working with us on proof-of-concept studies and pilot projects. To bring the technology into practical use, the natural path will be to develop it further through a spin-off. A central goal of the project is to ensure that this solution is accessible not only to large corporations but also to small and medium-sized enterprises.

Question: Looking ahead, do you see other potential “bridges” between physics, mathematics and AI making an impact on our everyday lives or the economy?

Nicola Serra: Yes. The boundary between physics, mathematics and AI is becoming increasingly productive. Particle physics has long dealt with enormous datasets, rare events, simulation-to-reality gaps, and the need for algorithms that behave reliably under uncertainty. These are exactly the challenges many industries now face. Methods such as generative modelling, robust optimisation, and graph neural networks have already proven powerful in areas like healthcare, transportation and energy. I expect these cross-disciplinary “bridges” to continue growing, because they allow us to bring mathematically solid, interpretable and scalable tools to domains where the stakes are very tangible.

A colleague once said (though I can’t trace the original source) that thermodynamics began as a tool to improve steam engines, and once its principles were understood, it helped enable technologies that eventually took humans to the Moon. It’s a good reminder that when mathematics and physics solve practical problems, they often open unexpected paths forward.

Barbara Warmbein