Tesseract Hackathon 2026
August 3–31, 2026 · online · solo or teams of up to 4
Compose a differentiable scientific workflow from multiple Tesseracts, and use end-to-end gradients to solve a real design, inference, or training problem.
Four weeks to build something that couldn't exist within a single framework. Compete for a share of the $20,000 prize pool and a head start on your next research paper.
The challenge
Use Tesseract to elevate differentiable systems#
Real scientific workflows span tools, languages, and machines, and that is exactly where gradients usually die. This year's challenge is to build a pipeline where they don't: multiple Tesseracts, one differentiable function, and an optimization, training, or inference loop that puts it to work.
Criteria
01 · compositionIt composes two or more Tesseracts across a real boundary
The components should disagree about how they compute, yet still compose into an end-to-end differentiable function. The boundary can be one of differentiation strategy, language, framework, or hardware. What matters is that it's a boundary that's otherwise hard to cross.
02 · gradientsThe gradients do the work
They drive a gradient-based optimizer, a training loop, or an inference scheme on a problem that can't be solved component by component, such as a coupled equilibrium, an inverse design, or a calibration. Show the objective improving and the recovered solution.
03 · framework fitThe workflow genuinely needs Tesseract
Your writeup should make the case for why this is a Tesseract problem: what was impossible or painful before (a language barrier, incompatible AD strategies, conflicting dependencies, a remote GPU, orchestration on HPC) that composition across the boundary resolves. We'd rather see a modest problem where Tesseract is clearly load-bearing than an ambitious one where it's a costume.
Two things worth knowing
Aim at a real problem
A prototype pointed at something a practitioner in the field would actually care about is how you win. For a sense of the bar, see last year's winning entries and the rocket-fin shape optimization case study that composes three Tesseracts across CAD, meshing, and an FEA solver.
Hack Tesseract if you need to
Hit a missing feature, a performance bottleneck, or a critical bug? Fork Tesseract and patch it yourself. We love to see teams that modify the software to fit their needs, as long as it's still recognizably Tesseract.
Tracks
Five ways in#
Pick the track that best describes your headline result. Cross-track submissions are fine; choose the track for the result you're proudest of.
Inverse design & shape optimization
A geometry/mesh/solver pipeline optimized end to end. Strongest entries cross a tool boundary (e.g. CAD or mesher ↔ solver) where the components differentiate differently: a finite-difference or surrogate-gradient mesher feeding an autodiff solver, an Enzyme-differentiated kernel composed with JAX, or an adjoint solver feeding autodiff post-processing.
Multi-physics & coupled systems
Two or more solvers from different domains coupled into one optimizable pipeline, with gradients flowing through the coupling. Ideally two-way, where each solver feeds back into the other and the optimum depends on their coupling.
Hybrid ML + mechanistic models
A learned component (closure, sub-grid model, surrogate, neural operator) trained by backpropagating through a physics solver that lives behind a Tesseract: the solver keeps its own container, language, and AD strategy, and gradients still flow from your loss, through the solver, into the network's weights. The interesting cases recover something physically meaningful, not just a black-box fit.
Differentiable inference & UQ
An expensive or black-box solver wrapped as a Tesseract and dropped into a probabilistic workflow (NumPyro, oryx, PyMC, an SBI scheme) for Bayesian calibration or an inverse problem. The solver may expose its Jacobian by autodiff or by finite differences; the composition with the inference engine is the contribution.
Differentiable graphics & rendering
A differentiable renderer, rasterizer, or geometry-processing component composed with a physics or ML Tesseract: inverse rendering, image-driven shape or material recovery, simulation-in-the-loop rendering. Gradients flow from a rendered image or geometric quantity back through the pipeline to the parameters that produced it.