In both pure mathematics and string theory, researchers frequently encounter vast, discrete landscapes of geometric objects. For instance, the Kreuzer-Skarke database contains over 473 million four-dimensional reflexive polyhedra, each acting as a combinatorial blueprint for constructing Calabi-Yau threefolds, as cataloged in A Calabi-Yau database: threefolds constructed from the Kreuzer-Skarke list. Similarly, searching for specific particle physics models within these spaces can yield solution counts on the order of $10^{723}$ (Counting string theory standard models).
Analyzing these spaces using traditional algebraic geometry is impossible due to their sheer scale. To find specific structures—such as manifolds that can stabilize the universe's volume or shapes that contain no internal lattice points—researchers must turn to algorithmic screening. The core insight of this field is that complex, continuous geometric properties can be mapped to discrete combinatorial invariants, allowing computers to rapidly filter out unviable candidates.
From Infinite Spaces to Finite Candidates
How do we search an infinite or astronomically large space? The key is to find a constructive combinatorial framing that bounds the search.
For example, consider the problem of classifying maximal lattice-free polyhedra—shapes that contain no integer points in their interior and cannot be expanded without capturing one. In Maximal lattice-free polyhedra: finiteness and an explicit description in dimension three, Gennadiy Averkov et al. proved that by generating candidates based on facet counts and using computational enumeration, an infinite search space can be reduced to a finite set of discrete combinatorial types that can be verified against specific lattice properties.
A similar logic applies when discovering new algebraic varieties. In Gorenstein Formats, Canonical and Calabi-Yau Threefolds, Gavin Brown et al. leveraged "Gorenstein formats" to systematically discover new families of Calabi-Yau threefolds. Instead of guessing equations blindly, the authors selected a format, enumerated valid weights for an ambient weighted projective space, and computed the Hilbert series. This turns a vague geometric search into a structured, deterministic pipeline.
This constructive approach is also vital in discrete geometry. When searching for specific point-set configurations, such as crossing families, the authors of Crossing and Non-Crossing Families designed constructive proofs that directly translate into efficient algorithms for isolating these families. Similarly, Elad Hazan et al. developed constructive algorithms in Volumetric Spanners: An Efficient Exploration Basis for Learning to build volumetric spanners for both discrete and continuous sets, making abstract geometric existence theorems computationally accessible.
Algorithmic Filtering for Physical Vacua
In string theory, finding a realistic model requires Calabi-Yau manifolds with highly specific internal structures. One prominent example is the Large Volume Scenario (LVS), which requires a "Swiss Cheese" geometry—a large overall volume dominated by a single big divisor, punctuated by several small holes (small 4-cycles).
To find these rare shapes among millions of possibilities, researchers build algorithmic tests that translate these physical requirements into scaling properties of geometric quantities. In Necessary conditions on Calabi-Yau manifolds for large volume vacua, James Gray et al. formulated an algorithmic test based on Kähler parameters and intersection numbers, applying it to Complete Intersection Calabi-Yau (CICY) datasets to reveal that a significant fraction of these geometries actually possess the necessary Swiss Cheese structure.
To scale this up, the authors of New large volume Calabi-Yau threefolds developed an algorithm to scan the massive Kreuzer-Skarke database directly. By analyzing the structure of reflexive polyhedra and searching for specific configurations of toric divisors, they systematically isolated thousands of new Swiss cheese geometries. This shifted the field from ad-hoc, lucky discoveries to a structured, exhaustive pipeline.
Bypassing the Bottlenecks with Machine Learning
Even with optimized combinatorial filters, certain geometric calculations—like computing exact triangulations or verifying topological fibrations—remain computationally bottlenecked. The latest evolution in this research direction replaces explicit algorithmic loops with statistical and reinforcement learning.
For instance, identifying whether a manifold possesses an elliptic fibration traditionally requires intensive algebraic calculations. In Distinguishing elliptic fibrations with AI, Seung-Joo Lee demonstrated that a neural network trained on CICY configuration matrices can accurately distinguish elliptically fibered manifolds from non-elliptically fibered ones, drastically outperforming traditional algebraic geometry methods.
Furthermore, navigating the combinatorial explosion of ways to triangulate a polytope can be framed as a game. In Generating Triangulations and Fibrations with Reinforcement Learning, Per Berglund et al. applied Deep Reinforcement Learning (RL) to train an agent to perform triangulations on complex reflexive polytopes from the Kreuzer-Skarke database. By rewarding the agent for finding configurations that satisfy strict mathematical and physical criteria, the RL agent learns to navigate these astronomically large landscapes far more efficiently than brute-force software.
Key Foundational Work
Necessary conditions on Calabi-Yau manifolds for large volume vacua (2012) — contributed 3 concept elements to this direction.
Maximal lattice-free polyhedra: finiteness and an explicit description in dimension three (2010) — contributed 1 concept elements to this direction.