Introduction
In the early 1990s, the landscape of AI-driven scientific discovery was characterized by a collection of highly successful but fragmented "special-purpose programs." Systems such as DALTON (chemistry), GELL-MANN (physics), and MENDEL (biology) demonstrated that computers could replicate historical discoveries, yet they appeared to operate on fundamentally different principles. This paper addresses the lack of a unitary core in these systems, proposing a unified mathematical framework that re-interprets scientific model-building as a search process within matrix spaces.
The Unified Matrix Framework
The authors propose that the process of postulating hidden structures, properties, and processes across diverse scientific domains can be standardized using three recurring matrix types:
- Reaction Matrix ($R$): Represents processes (e.g., chemical reactions or particle interactions) where rows correspond to reactions and columns to substances or entities.
- Structure Matrix ($S$): Describes the composition of entities, such as which atoms constitute a molecule or which quarks form a subatomic particle.
- Property Matrix ($P$): Assigns numeric values, such as electric charge or mass, to fundamental entities.
By representing scientific knowledge in this way, the "search" for a new model becomes a search for the correct values to fill these matrices subject to domain-specific constraints.
Conservation Laws as Algebraic Constraints
A central insight of this framework is that fundamental conservation laws—such as the conservation of mass in chemistry or charge in physics—can be expressed as simple matrix equations. For instance, the requirement that atoms be conserved across a chemical reaction is captured by the equation $R \times S = 0$.
As shown in the example above, the stoichiometric coefficients of a reaction (the $R$ matrix) and the atomic composition of the substances (the $S$ matrix) must satisfy a null-space relationship. This mathematical identity allows discovery systems to move away from slow, symbolic search and instead utilize efficient linear algebra tools like Gaussian elimination and linear programming.
Model-Building as Matrix Search
When a current model fails to satisfy observed data, the system "searches" by progressively enlarging the dimensions of the matrices. Adding a new column to a structure matrix corresponds to postulating a new hidden entity, such as a new quark or a chemical element. This approach naturally incorporates Occam’s Razor: the system starts with the smallest possible matrices and only increases complexity when the current size cannot satisfy the constraints.
This generalization allows for the integration of disparate fields. For example, the task of finding quarks (GELL-MANN) and discovering conserved quantum properties (PAULI) can be merged into a single matrix constraint system ($S \times P = P'$), where $P'$ represents the observed properties of the composite particles.
Conclusion
By providing a formal re-interpretation of six major discovery systems—ranging from qualitative chemical models to quantitative particle physics—this paper demonstrates that a "universal discovery engine" is theoretically possible. While some challenges remain in integrating graph-based models (like molecular structures) and purely descriptive empirical laws, the matrix-space framework offers a powerful, generalized account of how scientific models are constructed and validated.
Summary updated