Alain Richardt, founder of Atomic Tessellator and former Google engineer, walks through the architecture of a materials discovery platform that treats physics simulation as the foundation, not an afterthought. The core loop: run density functional theory (DFT) on tens of thousands of candidate crystal structures, use those results to train a surrogate model that approximates the physics at lower computational cost, then apply active learning to select the next batch of candidates for high-fidelity simulation. This isn’t generative chemistry throwing darts at a latent space—it’s a deliberate, constraint-driven search over structures that satisfy specific property targets like magnetic anisotropy, thermal stability, or dielectric breakdown strength.
The conversation gets granular fast. Richardt explains how they parameterize the search space using symmetry-aware representations that encode crystallographic group information, site occupancy, and lattice strain—representations that respect the underlying physics rather than flattening everything into a generic graph neural network input. He walks through a case study on rare-earth-free permanent magnets, where the platform screened for compounds with high saturation magnetization and uniaxial anisotropy, ultimately surfacing a candidate that matched early neodymium-iron-boron performance without a single rare-earth element. The timeline: weeks from problem definition to candidate identification, versus the typical multi-year, Edisonian trial-and-error cycle.
On fusion materials, Richardt details the multi-objective optimization challenge: a first-wall material must simultaneously tolerate neutron damage, resist sputtering, and maintain thermal conductivity under steady-state heat flux. No single property dominates, and the trade-off surface is non-convex. Their approach uses Pareto-front optimization across DFT-calculated property landscapes, then validates top candidates with molecular dynamics simulations of cascade damage. The result is a ranked list of synthesizable candidates with predicted operational lifetimes—actionable output for reactor engineering teams.
The semiconductor segment focuses on a problem that’s quietly becoming critical: as gate oxides scale below 1 nm, defect densities that were once negligible become dominant failure modes. Atomic Tessellator’s models predict vacancy formation energies and migration barriers across thousands of candidate dielectrics, identifying materials where oxygen vacancies are both hard to form and slow to move. Richardt notes that several candidates they’ve surfaced are not in any existing materials database—they occupy entirely unexplored regions of phase space.
Throughout the episode, Richardt emphasizes that the real bottleneck isn’t compute or algorithms—it’s the experiment selection problem. With ~10^60 possible inorganic crystals and lab synthesis costing thousands of dollars per attempt, the value is in knowing which 10 compounds to actually make. Their active learning framework explicitly optimizes for information gain per synthesis dollar, and he shares data showing that their approach reduces the number of experiments required to find a target material by two orders of magnitude compared to high-throughput screening alone. For anyone building at the hardware frontier, this episode is a masterclass in how to make AI actually useful for physical discovery—not by replacing physics, but by using it as the foundation.
Key Insights
- Atomic Tessellator runs DFT simulations on 10^4–10^5 candidate structures before training a surrogate model, inverting the standard ML pipeline so predictions are constrained by physics, not just historical data.
- Their AI discovered a rare-earth-free magnet candidate in weeks that matched the energy product of early neodymium magnets—without touching a lab, by screening crystal structures that maximize magnetic anisotropy and saturation magnetization.
- For fusion reactor materials, the platform screens refractory alloys and ceramics against simultaneous neutron flux, thermal load, and sputtering constraints—a multi-objective optimization problem that manual trial-and-error cannot solve in realistic timeframes.
- The company’s semiconductor work focuses on dielectric materials with ultra-low defect densities; their models predict vacancy formation energies and migration barriers to identify candidates that resist breakdown under high-k scaling.
- Richardt argues that the bottleneck in materials discovery is not compute but the 'experiment selection problem'—knowing which 10 compounds to synthesize out of 10^60 possible inorganic crystals—and their active learning loop directly addresses this.
- The platform generates not just candidate materials but the synthesis pathway and predicted stability window, reducing the translation gap between computational prediction and experimental realization that kills most academic discoveries.
Who should listen: Computational materials scientists, semiconductor process engineers, and deep tech founders building hardware products where materials performance is the binding constraint on system-level innovation.
Why This Matters
This episode operationalizes the 'simulation-first' thesis we track at the frontier: rather than treating AI as a pattern-matching black box, Atomic Tessellator uses DFT as a ground-truth oracle and lets ML interpolate between known physics. It’s a concrete case study in how the next generation of deep tech companies are rebuilding the discovery stack from first principles—with implications for anyone working on hardware-software co-design, advanced manufacturing, or energy systems where materials are the rate-limiter.
