Scientific Intelligence Substrate

• Scientific Intelligence Substrate is a framework of physical, chemical, computational, and organizational systems that shape scientific reasoning and discovery.
• It integrates molecular mechanisms, unconventional chemical media, and engineered infrastructures to support adaptive, high-fidelity intelligence.
• Research leverages quantitative metrics, such as fractal hydropathic profiling and information-theoretic measures, to validate substrate efficacy.

A scientific intelligence substrate is any physical, chemical, computational, or organizational platform whose microstructure and organizational principles enable the emergence, expression, or enhancement of scientific reasoning, learning, and discovery. Such substrates are not mere media or implementation details but actively shape the cognitive, computational, or physical characteristics of intelligence at all levels, from molecular to organizational. Across biological, artificial, and hybrid settings, the term encompasses minimally sufficient neural circuits, specialized protein geometries, distributed physical or computational networks, and engineered or federated infrastructure for autonomous research. Substrate properties—molecular, informational, architectural, dynamical—thus impose constraints and affordances on the expression of intelligence and its scientific applications.

1. Molecular and Biophysical Substrates

At the molecular scale, intelligence substrates can be traced to biochemical and biophysical architectures influencing neural computation. Kif14, a kinesin motor protein central to axonal vesicle transport, exhibits evolutionarily acquired synchrony in hydrophobic extrema of its amino acid sequence, quantified by water-wave hydropathic profiling using a fractal-based scale. For 12 vertebrate species, the degree of leveling in the top five hydrophobic peaks (mean deviation MD_s) is minimized in Homo sapiens (1.50) and increases monotonically down the phylogenetic tree (up to 4.10 for naked mole rat), indicating >2.5× sharpening in human Kif14 (Phillips, 2021).

Mathematically, solvent-accessible surface areas (SASA) are analyzed as

$$\log \mathrm{SASA}(L) \simeq \text{const} - Y(\mathrm{aa}) \log L$$

for segment lengths $9 \leq L \leq 35$, defining a universal fractal hydropathic index $\psi_\mathrm{fractal}(i) = Y(\mathrm{aa}_i)$. When convolved over a window of optimal width $W^* = 9$, leveled hydrophobic “peaks” are interpreted as synchronized domain motion barriers in the kinesin motor. This synchronization narrows dwell-time distributions for ATP-coupled stepping and supports precise, jitter-minimized delivery of synaptic vesicles. The proposed consequence at the neural network level is improved synchrony and dynamic range—core attributes of adaptive, high-fidelity neural computation. Thus, substrate-level molecular geometry can be evolutionarily honed to underpin key biophysical features of higher-order intelligence (Phillips, 2021).

2. Substrate Principles in Artificial and Unconventional Media

Classic and unconventional substrates for intelligence encompass both algorithmic and physical domains. Turing’s unorganised machines—a class of discrete dynamical systems made from randomly interlinked recurrent NAND gates—have been experimentally implemented in compartmentalized excitable chemical media, such as networks of Belousov–Zhabotinsky (BZ) vesicles (Bull et al., 2012). Each vesicle acts as a processing node, with excitation dynamics governed by reaction–diffusion equations (Oregonator or FitzHugh–Nagumo models):

$$\frac{\partial X}{\partial t} = \frac{1}{\varepsilon}(X - X^2 - (fY + \phi) X) + D_X \nabla^2 X$$

Logical gates, including NAND, XOR, XNOR, are built from collision-based geometries, and network structure is discovered via imitation programming. Successful realization of benchmark logic circuits demonstrates that complex, recurrent computation and thus substrate-level intelligence can be chemically instantiated, not merely simulated digitally.

This result supports the notion that intelligence substrates are not restricted to canonical neural or silicon architectures but can be constituted by diverse media—so long as they support the requisite dynamical and combinatorial complexity for computation and adaptation (Bull et al., 2012).

3. Information-Theoretic and Organizational Perspectives

Intelligence as an emergent property of substrate organization is formalized in information-theoretic terms (Gershenson, 2021). Regardless of physical realization (neural tissue, swarms, computers), the substrate’s capacity to process, integrate, and transmit information underlies measurable intelligence. Core organizational metrics include:

• Shannon entropy $H(X)$: quantifies surprise or information content.
• Mutual information $I(X;Y)$: measures the reduction in uncertainty in $X$ from knowing $Y$.
• Integrated information $\Phi$: quantifies the extent to which the whole system’s dynamics cannot be reduced to independent parts.
• Algorithmic (Kolmogorov) complexity $K(s)$: the shortest program generating $s$.

In biological brains, high $\Phi$ and network complexity are correlated with cognitive integration and adaptive behavior; in synthetic substrates such as computer networks or swarms, intelligence emerges when organizational complexity enables robust, goal-driven transformation of information.

This strongly substrate-neutral stance is echoed in the Theory of Intelligences (TIS) (Hochberg, 2023), which posits macroscopic, informational equations to quantify problem solving and planning across physical, biological, and computational substrates. Solving ability $\mathbb{U}_y$ and planning efficiency $\mathbb{A}_y$ are combined to yield an intelligence index $\mathbb{I}_{x,y}$:

$$\mathbb{I}_{x,y} = \frac{1}{2} \mathbb{U}_{x,y}^R + \frac{1}{2}\mathbb{U}_{x,y}^R \mathbb{A}_{x,y}^G,$$

where quantities are normalized relative to capacity and problem complexity. Proxies, such as technology or group knowledge, multiply substrate abilities—formalizing evolutionary and cultural bootstrapping of scientific intelligence (Hochberg, 2023).

4. Scientific Intelligence Substrate in AI and Scientific Systems

In engineered and computational intelligence, substrates are realized as multi-layered architectures that integrate data, models, memory, and domain-specific knowledge in a single platform (Wolff, 2013, Heredia et al., 18 Dec 2025, Zhang et al., 24 Jun 2025). The SP Theory of Intelligence (Wolff, 2013, Wolff, 2018) proposes a minimal, Turing-complete substrate grounded in information compression through multiple alignment of atomic symbol patterns. Every cognitive function—recognition, reasoning, planning, learning—emerges from alignment and compression, rather than from separated symbolic or numeric modules.

This substrate is realized concretely as a system storing both “New” and “Old” patterns, and searching for alignments maximizing compression:

$\mathrm{CD} = B_N - B_E$

where $B_N$ is the number of bits encoded from New patterns, and $B_E$ is the bit cost of the code pattern. All inferences, including probabilistic and nonmonotonic reasoning, arise from this central compression mechanism. Learning proceeds by constructing grammars that minimize the combined description length of knowledge base and compressed codes.

In LLM-based scientific systems, the substrate expands to encompass planners, memory, and toolsets interconnected in a modular pipeline. Knowledge is integrated via retrieval-augmented models, structured knowledge graphs, and fine-tuning objectives, while tool interfaces connect symbolic reasoning to executable scientific domains (Ren et al., 31 Mar 2025). Key mathematical learning objectives—supervised fine-tuning loss

$$\mathcal{L}_\mathrm{SFT}(\theta) = -\mathbb{E}_{(x, y)} \sum_{t=1}^T \log P_{\theta}(y_t|x, y_{<t})$$

and RL with KL regularization

$$J(\theta)= \mathbb{E}_{y\sim \pi_\theta}[r(y,x)] - \lambda D_{\mathrm{KL}}(\pi_\theta || \pi_\mathrm{SFT})$$

formalize the substrate’s adaptivity.

Federated infrastructures (e.g., AI4EOSC) generalize the substrate concept to distributed, multi-institutional environments—providing unified access to computation, storage, training, and model deployment, while ensuring traceability (W3C PROV), reproducibility, and secure multi-organizational federation (Heredia et al., 18 Dec 2025). Such platforms operationalize scientific intelligence as a system-level emergent property built atop technical, procedural, and organizational substrate layers.

5. Minimal and Sufficient Substrate Conditions

Minimality is empirically addressed by evaluating neural-network models with constrained capacity and task-general training. The LaMa network, a feed-forward convolutional model with $\sim$51M parameters and Fast-Fourier Convolutional modules, achieves human-equivalent scores on Raven’s Progressive Matrices—a canonical test of fluid intelligence—despite being drastically smaller than the human brain and trained solely on natural scene inpainting (Nelson et al., 2023).

This result establishes that “substrate validity”—the minimal configurations (structural, computational) required for specific scientific intelligence features—is both quantifiable and far less demanding than biological precedent. Computationally simple, non-symbolic substrates can suffice for complex cognitive inference, provided they have sufficient global pattern integration capacity.

6. Substrate Dependence, Physicalism, and Consciousness

The relationship between substrate properties and higher-order phenomena such as consciousness and scientific reasoning is formalized in models that distinguish between the reasoning process ($R$) and the underlying physical substrate ($P$), with critical dependence on substrate features such as autonomous reactive subsystems (Sritriratanarak et al., 17 Oct 2025). The necessity for “consciousness switches,” denoted $c$, is shown to depend on the presence of reactive, autonomous responses ($a_r$) in $P$:

• If $P$ is reactive, $R$ must possess $c$ to avoid harmful “real” actions triggered by simulated or hypothetical thinking steps.
• If $P$ is non-reactive, intelligence (reasoning depth, learning) becomes fully substrate-independent and does not require consciousness.

Thus, the necessity and nature of higher faculties are substrate contingent, and the addition of such faculties (e.g., consciousness) does not enhance intelligence per se except by protecting the substrate from “damage by thinking.” The model confirms physicalism: all phenomena are ultimately computationally and physically grounded (Sritriratanarak et al., 17 Oct 2025).

7. Scientific Intelligence Substrate in Experimental and Collaborative Contexts

Intelligence substrates in experimental science are realized as integrated, closed-loop systems combining cognitive, agentic, and embodied capabilities (Zhang et al., 24 Jun 2025, Lin et al., 3 Nov 2025, Schmidt et al., 20 Feb 2024). Architectures such as the Intelligent Science Laboratory (ISL) employ foundation models for reasoning, agent-based workflow orchestration, and embodied experimentation agents in a feedback loop. Closed-loop adaptation, sim-to-real policy transfer, and digital twins are canonical substrate features, supporting continual experimental refinement and serendipitous discovery.

Human-AI co-embodiment systems (e.g., APEX) unify human operators, agentic AI, and wearable interfaces in a composite substrate $\langle S, A, H, P, R, M, F\rangle$, delivering real-time perceptual analysis, multimodal memory, adaptive planning, and high-fidelity, traceable protocol execution. Quantitative improvements exceeding 50% in task fidelity, error correction, and knowledge transfer demonstrate that such substrates fundamentally alter the landscape of scientific research and manufacturing, uniting interpretability, scalability, and continuous adaptation with physical action (Lin et al., 3 Nov 2025).

• "Kinesin Motors and the Evolution of Intelligence" (Phillips, 2021)
• "The SP theory of intelligence: an overview" (Wolff, 2013)
• "Consciousness, natural and artificial: an evolutionary advantage for reasoning on reactive substrates" (Sritriratanarak et al., 17 Oct 2025)
• "Position: Intelligent Science Laboratory Requires the Integration of Cognitive and Embodied AI" (Zhang et al., 24 Jun 2025)
• "The minimal computational substrate of fluid intelligence" (Nelson et al., 2023)
• "Toward Turing's A-type Unorganised Machines in an Unconventional Substrate" (Bull et al., 2012)
• "Intelligence as information processing: brains, swarms, and computers" (Gershenson, 2021)
• "The SP theory of intelligence and the representation and processing of knowledge in the brain" (Wolff, 2016)
• "Virtual Reality for Understanding Artificial-Intelligence-driven Scientific Discovery" (Schmidt et al., 20 Feb 2024)
• "AI4EOSC: a Federated Cloud Platform for Artificial Intelligence in Scientific Research" (Heredia et al., 18 Dec 2025)
• "Towards Scientific Intelligence: A Survey of LLM-based Scientific Agents" (Ren et al., 31 Mar 2025)
• "Human-AI Co-Embodied Intelligence for Scientific Experimentation and Manufacturing" (Lin et al., 3 Nov 2025)
• "Introduction to the SP theory of intelligence" (Wolff, 2018)
• "AI4X Roadmap: Artificial Intelligence for the advancement of scientific pursuit and its future directions" (Dale et al., 26 Nov 2025)
• "A Theory of Intelligences" (Hochberg, 2023)
• "Probing Scientific General Intelligence of LLMs with Scientist-Aligned Workflows" (Xu et al., 18 Dec 2025)