Japan's New Algorithm Finds Crystal Structures in Seconds on a Laptop

The crystallography scandal just got cheaper

For nearly forty years, crystal structure prediction has been called a scandal in physics. The problem is embarrassingly simple to state — given a set of elements, figure out how their atoms arrange themselves in three dimensions — and impossibly hard to solve in practice. You need to search through an astronomically large space of possible atomic positions and periodic arrangements, then evaluate each candidate with quantum-mechanical first-principles calculations. Even supercomputers struggle.

A team at the University of Tokyo has found a way around the bottleneck by refusing to do the expensive calculation until absolutely necessary. The result, published in Physical Review Research, turns crystal structure prediction from a supercomputing problem into something you can run on a laptop in seconds — for simple structures at least.

The lead researcher is Ryotaro Koshoji, a specialist appointee at the Institute for Solid State Physics. The work extends a 2024 framework he called “mathematical crystal chemistry” into a complete theory. The key insight is surprisingly old-fashioned: inorganic chemists have long described crystals using two intuitive rules — atoms pack like spheres, and they also form networks of coordination polyhedra. Neither description alone is sufficient, but together they constrain the solution space so tightly that most unstable configurations simply vanish.

Math as a filter, not a simulator

Koshoji’s team formulated those empirical rules as a generalized disjunctive programming problem — a type of mixed-integer nonlinear optimization that handles both continuous variables (atomic positions) and discrete variables (which bonds exist and which don’t). The result is a mathematical filter. Feed it atomic radii and coordination numbers, and it systematically eliminates candidates that violate basic geometric and bonding constraints. What remains are structures worth taking seriously.

The algorithm blends several optimization techniques — logic-based outer approximation, branch-and-bound, cutting-plane methods, and machine learning — into a single pipeline tuned to crystallography. In tests, the team showed it could predict ionic crystal structures and Zintl phases (materials containing both ionic and covalent bonds). They also demonstrated exploratory searches on quaternary oxide systems, designing structures that don’t correspond to any known experimental compound.

The software, named MARICI, is released under an MIT license with a Windows GUI. Simple structures take seconds. Complex ones still benefit enormously — the team reported tenfold speedups over conventional approaches — but the real story is what gets excluded before any quantum calculation runs.

Why this matters beyond the journal

The deeper significance is structural. Most crystal structure prediction methods generate candidate structures randomly or with genetic algorithms, then test stability with expensive calculations. They reduce the search space but still require quantum mechanics for every promising candidate. Koshoji’s approach flips the order: use cheap mathematics to generate only “crystal-like” structures, then reserve first-principles calculations for verification. The quantum work becomes the final step, not the engine.

This matters because the materials pipeline for batteries, semiconductors, and pharmaceuticals is constrained not by lack of ideas but by the cost of evaluation. A new electrolyte for solid-state batteries might exist in a compositional space that has never been computationally explored because the structure prediction step alone requires weeks on a cluster. Remove that wall, and the exploration frontier expands dramatically.

Japan has a pattern of producing computational materials science tools that outperform Western alternatives precisely because they combine deep domain knowledge with rigorous mathematics rather than brute-force computation. The Fujitsu supercomputer Fugaku has been used extensively for materials screening, but this new work is different — it doesn’t need Fugaku for the initial prediction phase. That’s the shift.

Who wins, who loses

The winners are anyone doing high-throughput materials discovery. Battery researchers hunting for new ionic conductors, semiconductor companies exploring wide-bandgap oxides, pharmaceutical firms studying polymorphs — all of them now have a tool that can pre-filter candidate structures at near-zero cost. The MIT license means no licensing barrier. A graduate student in Nairobi or São Paulo can run MARICI on a consumer laptop the same way a researcher at a Japanese national lab can.

The losers are the old workflow. Methods that rely on generating thousands of random structures and filtering them with density functional theory will look expensive by comparison. Not obsolete — DFT is still necessary for final stability verification — but no longer the first step.

There are limits the team acknowledges. The current framework handles ionic and Zintl-phase materials well; fully covalent or metallic systems remain harder. The team plans to develop symmetry-based initial structure generation and determine the minimal set of input conditions needed to describe most ionic crystals. If those pieces fall into place, the group expects large-scale computational screening of functional materials to become routine within a few years.

The wider point

What makes this story worth watching isn’t a single paper. It’s the trajectory. Japan’s Institute for Solid State Physics has been producing crystal chemistry tools for decades, and each iteration has pushed the computational boundary further while reducing the resource requirement. The 2024 version needed a mathematical framework. This version needs only a laptop and an MIT license.

The implication for global supply chains is indirect but real. Materials discovery is the first link in the chain from element to product. Speed that link up, and everything downstream — battery production, chip fabrication, drug development — moves faster. Japan doesn’t always lead in headlines about AI or consumer electronics, but in the quiet infrastructure of computational science, it keeps building tools that other people’s industries depend on.

The crystallography scandal isn’t over yet. But the pen has finally gotten cheaper.