SYREQ™ uses symbolic regression to derive compact equations from industrial data, compute-intensive simulations, and existing models. Rather than fitting a predetermined form, it searches for the variables, operations, structure, and values that best describe the system.
Fast to evaluate. Use compact equations repeatedly inside optimization and control systems.
Open to inspection. Compare each equation’s structure with physical knowledge, test constraints, and determine where each model is valid.
Create a model from measured process or system data when no suitable mathematical form is known in advance. Domain experts can inspect which variables matter, how they interact, and whether the relationship is physically plausible.
Approximate simulations and other compute-intensive models with compact surrogate equations. Fast evaluation makes repeated calls practical in optimization, control, parameter studies, and scenario analysis without running the source model each time.
Reconstruct an opaque, difficult-to-maintain, or hard-to-deploy model as an explicit equation. The replacement can be validated against the original, reviewed by domain experts, ported, and embedded in current systems.
A brief look at how SYREQ searches, optimizes, and selects equations.
SYREQ represents equations as expression trees and generates candidate structures from a bounded mathematical language. Exhaustive search covers that space systematically and retains a frontier showing how predictive accuracy changes as equations become more complex.
For every candidate structure, a numerical optimizer tunes its coefficients and constants against the data before the equation is scored.
Define which variables, operations, and structural patterns are meaningful. Because these rules apply while candidates are generated, domain knowledge shapes the search itself and focuses compute on plausible model families.
The model language is formally defined by a context-free grammar. It can express anything from a few operator restrictions to complete structural templates. The fixed branch shown here is only one example.
Exhaustive search covers a bounded model space methodically. When full coverage is impractical, genetic programming evolves valid equations through mutation and crossover, concentrating the search on promising regions while obeying the same grammar.
Genetic programming is currently in development.
Symbolic regression is computationally demanding. GPU acceleration makes large searches practical, while distributed execution extends the same approach across additional compute as model spaces grow.
Generate, fit, and evaluate large batches of candidate equations concurrently on GPU hardware.
Extend exhaustive and evolutionary searches across multiple workers as the model space grows.
Start with measured data, a compute-intensive simulation, or a model that no longer fits its next use. We work with industrial teams to define the target behavior, admissible model space, validation tests, and deployment requirements.
