The Atlas project is an independent research system designed to discover mathematical patterns and scientific laws by operating within a multidimensional space of interconnected knowledge. Unlike traditional neural networks that generate answers or fit formulas, Atlas functions as a navigational mechanism that allows new axes, operators, and concepts to emerge during the research process.
The system organizes knowledge into six interconnected components: observations and experimental data, scientific axes, candidate hypotheses, proofs and uncertainty intervals, research history, and provenance. As of October 2026, the system includes 13 scientific domains, 655 canonical axes, 445 relationship owners, and 24 computational methods, as documented in its mathematical books.
Atlas has progressed from identifying mathematical invariants to constructing unknown representations and testing hypotheses against physical measurements. In one study, the system identified a need for an additional mathematical coordinate to describe differences in vortex intensity between two fluid states that appeared identical under standard integral characteristics.
The project utilizes a neurodynamic mechanism to determine the most effective research actions, such as applying polynomial searches or checking symmetry, rather than relying on neural weights for truth. In tests using the SC17 build, the system demonstrated a 25% efficiency gain in finding first integrals across 24 mathematical tasks compared to structural strategies.
Atlas is being applied to complex scenarios involving multiple sensors and physical variables, such as analyzing aerodynamic resistance in racing cars or optimizing chip design. For microchip development, the system uses the EDA-CHIP-DESIGN-RESEARCH/1.0.0 module, a test project on the sky130hd platform, and integrates with the OpenROAD Flow.
