Continuous Physics Reasoning: Always-On Design Infrastructure
Interview with
Hardik Kabaria, Cofounder & CEO, Vinci
Hardik Kabaria returns to CDFAM in DC as cofounder and CEO of Vinci, a frontier lab building what the company calls continuous physics infrastructure.
Vinci has built a physics foundation model that computes thermal and thermo-elastic behavior directly on manufacturing geometry, continuously, as a design evolves, rather than as a discrete checkpoint engineers wait at for hours or days.
In the following interview, Hardik walks through the foundation model and the three physics it covers today, why the same approach generalizes to automotive, aerospace, energy, and medical devices, how Vinci ingests native design formats without simplification or manual meshing, and what it takes to move physics evaluation from a late checkpoint to something that runs inside the design loop from the first floorplan onward.
Can you start by telling us about Vinci and what the platform offers, and what your role there involves?
I’m a cofounder and CEO of Vinci. My background sits right at the intersection the company is built on: I led software at Carbon, and my Ph.D. is in mechanical engineering from Stanford. Large-scale software on one side, computational mechanics on the other, which is exactly the problem Vinci takes on.
Vinci is a frontier lab. We’re building what we call continuous physics infrastructure, the systems that make physical reality continuously computable.
In plain terms: today, simulation is a discrete step. You freeze a design, hand it to a specialist, and wait hours or days for an answer. We’ve built a physics foundation model that computes physical behavior directly on manufacturing geometry, continuously, as the design evolves.
Physics stops being a gate you wait at and becomes an always-on constraint. That’s what lets teams converge faster on manufacturable, high-yield designs. Today the model does three physics: static thermal conduction, transient conduction, and thermo-elastic warpage. And it works on new designs out of the box, with no per-customer training.
Vinci came out of the gate focused on semiconductor design. What drove that decision, and how do you see the approach extending into other industries?
We started with semiconductors on purpose, because it’s the hardest case. Advanced packaging has nanometer features sitting right next to centimeter-scale assemblies, with tightly coupled thermal, mechanical, and electrical effects, and correctness isn’t a nice-to-have, it’s a contractual sign-off requirement. The IP constraints are about as strict as they get, too. So if the model reasons reliably there, out of the box, that tells you the physical understanding is real, not pattern-matching to designs it has seen before.
And here’s the key point: physics doesn’t change across industries. Heat conduction in a chiplet stack is the same physics as in a battery pack or a power inverter. What changes is the geometry, the materials, and the scale. So the capability that clears the semiconductor bar carries straight over to automotive, aerospace, energy, and medical devices. We prove it where it’s hardest, then generalize outward, not the other way around.
Can you walk us through the foundation model: what physics domains it covers today, what is on the roadmap, and where the current limits of generalization sit?
Today it’s three physics in a single model, not three separate tools: static thermal conduction, transient conduction, and thermo-elasticity, which is the warpage problem.
The important part is what kind of model it is. It’s deterministic, it’s verifiable against first-principles solvers at manufacturing resolution, and it generalizes to new geometry and materials out of the box, with no per-customer training and no per-domain forks.
That’s a deliberate line we draw: a lot of physics AI is a trained surrogate, fast inside the design space it was trained on, but it has to be retrained or tuned for each new case. We don’t do that. In our published results, the model holds sub-percent error on geometries it has never seen, including out-of-distribution cases, in seconds where FEM takes hours.
The roadmap expands two ways. One is more physics, all of it governed by partial differential equations: convection and radiation to complete the heat-transfer picture, then nonlinear elasticity and signal integrity, and further out the harder regimes like RF and Maxwell’s equations, plasticity, and turbulent flow.
The other is up the value chain: from physics intelligence today, to systematically improving designs against design and manufacturing rules, to eventually generating manufacturing-ready designs from specification.
On the limits, I’m careful here. We generalize within the model’s declared scope, meaning the physics it covers today. Convection, for example, is on the near-term roadmap but it’s not in the model yet, so the fluid side enters as a boundary condition today, not a solved field. That discipline is deliberate. It’s how you tell a real foundation model from one that overclaims.
For engineers evaluating the platform, the practical question is always about integration. What does Vinci take as inputs, how does it connect to existing toolchains, and what does the output look like?
The primary path is native design data. The solver ingests the formats teams already work in: GDS and OASIS for layout, STEP and STL for 3D geometry, IPC-2581 and ECXML. There’s no conversion step and no rebuilding the model in a separate tool.
For early-stage work, when a design isn’t yet in one of those formats, the geometry can instead be defined through a structured configuration file that specifies the layer stackup, material assignments, and component bounding boxes. Either way, power is applied as a separate step depending on where in the package it needs to go, which keeps the pipeline reproducible and fully automated.
What makes the integration meaningful is that we ingest that data with no scale or resolution limitations. Vinci resolves BEOL metal layers, micro-bump arrays, and through-silicon vias at their actual dimensions, from nanometers to microns, with no simplification and no manual meshing. That preserves the vertical conduction through the vias and micro-bumps and the lateral spreading through the metal layers that homogenized models simply can’t capture.
The outputs are the field results engineers already expect: temperature fields, warpage and displacement, and time-resolved junction-temperature histories. They’re verifiable against first-principles solvers, and they feed the same sign-off decisions teams already make, just produced earlier in the design process.
The abstract for your presentation argues physics evaluation belongs earlier in design. For a team running traditional FEM-based workflows, what does that transition actually look like in practice?
First thing I’d say: it’s not “throw out FEM.” FEM stays. It’s your ground truth, your verification instrument. What changes is the order of operations.
In a traditional flow, physics is a checkpoint. You freeze a candidate, mesh it, set up the solve, wait hours, interpret the result, and hand back feedback, and by then the design has often moved on. So you only ever check a handful of candidates, and usually only late, when changing anything is expensive.
What unlocks the shift is that the whole workflow gets faster, not just the solve. Because we work directly from geometry with no simplification, with automated meshing and an AI-accelerated solver, the end-to-end run is on the order of a thousand times faster than traditional FEA. When an answer comes back in that timeframe with no manual setup, you stop treating physics as a gate and start putting it inside the loop. Instead of checking one design at the end, you can run a design-of-experiments across hundreds of points: sweeping power maps, TIM thickness, TIM conductivity, and floorplan and cooling combinations, and find the ones that buy you thermal margin at the same cost.
And to be clear, this doesn’t take the engineer out of the loop. It expands what they can do. The engineer still defines the constraints, interprets the results, and makes the final call.
What changes is that they’re no longer waiting on simulation runs or rationing how many designs they can afford to check. Their role shifts from running individual solves to asking the right questions and validating the answers. Physics goes from a gate the design has to pass to a partner that shapes it from the first floorplan onward.
What do you hope to take away from participating at CDFAM DC, and what is the one thing you want the audience to leave with?
What I want from CD/DC is the cross-section it pulls together: engineering leaders, software developers, suppliers, and government, all pushing to get computational and AI methods into design earlier. I especially want the hard questions from people outside semiconductors, because that’s exactly how the generalization argument gets stronger.
Where I want this to go is bigger than any one industry. The same physics that gates an advanced package also gates a battery pack, an aircraft structure, an implantable device, a power grid component. In a sense it’s a democratization of physics: today, only a small number of specialists can answer these questions, and we want to put that capability in the hands of far more engineers.
The goal is that designing against real physics from the first concept becomes the default, not a privilege of the few teams that can afford long simulation cycles. Semiconductors are where we prove it. From there, the aim is to bring the same capability to every industry building physical products at scale.
The one thing I want people to leave with: this is not faster simulation, it’s a different operational mode.
When solver-grade physics is continuously available instead of an occasional checkpoint, physics stops being an end-of-cycle gate and becomes always-on design infrastructure. And that changes not just how fast you design, but what you can design at all.
CD/DC brings together the people working to get computational and AI methods into design earlier, across both industry and government. Join us in Washington DC on July 15-16, 2026 to meet Kabaria and others working on the same problems in person.
