Pillar I · The scientific learner
Foundational Scientific Learning-System Design
We design scientific representations, model architectures and information flow, learning signals and objectives, and the ways models learn and adapt. These choices determine which distinctions a system can preserve, which relationships it can express, and how it can use limited evidence.
We also study how models, data, simulation, experiments, tools, and scientific judgment work together as a complete system. Research infrastructure and scientific evaluation are part of this design: they shape what can be tested, compared, and reused, and how progress toward scientific understanding and reliable design is assessed.
Explore the learner
Keep the four input values fixed. Change which units interact and see what each calculation can preserve.
Illustrative arithmetic, not measured data, a trained model or a physical interaction.
Ignores the connections; stays the same for all three pairings.
Uses the connections; distinguishes these three pairings.
(1 × 2) + (3 × 4) = 14
What a learner can distinguish depends on which information its representation and computations retain. The inputs alone do not specify the relationships.
Ignores the connections; stays the same for all three pairings.
Uses the connections; distinguishes these three pairings.
(1 × 3) + (2 × 4) = 11
What a learner can distinguish depends on which information its representation and computations retain. The inputs alone do not specify the relationships.
Ignores the connections; stays the same for all three pairings.
Uses the connections; distinguishes these three pairings.
(1 × 4) + (2 × 3) = 10
What a learner can distinguish depends on which information its representation and computations retain. The inputs alone do not specify the relationships.
1–2 and 3–4
Mean of raw features: 2.5
(1 × 2) + (3 × 4) = 14
1–3 and 2–4
Mean of raw features: 2.5
(1 × 3) + (2 × 4) = 11
1–4 and 2–3
Mean of raw features: 2.5
(1 × 4) + (2 × 3) = 10
What a learner can distinguish depends on which information its representation and computations retain. The inputs alone do not specify the relationships.
This calculation preserves the distinction between these three pairings. It does not distinguish every possible graph or establish a physical mechanism. Pooling after an interaction-aware update can retain information that averaging raw inputs loses. A useful scientific learner also needs appropriate representations, guiding signals and learning dynamics.
A representation must preserve the distinctions the scientific question depends on. In this small example, the feature values stay fixed while the connections change.
Each numbered node is an information unit carrying one scalar feature. Lines specify pairs in the example; their position and length have no physical meaning.
The example motivates designing which units interact and how their information is combined. It does not show that every graph can be distinguished this way, or that pooling after an interaction-aware update necessarily loses this information.
View full-size diagram: Wide layout Vertical layout
Representations organise scientific states; interactions connect them; learning revises the system using evidence and guiding signals. These choices are designed together. This conceptual illustration does not specify a model architecture; the worked example on this page makes one interaction choice explicit.
Conceptual framework
View full-size framework: Wide layout Vertical layout
Define the distinction, mechanism, or decision the learner must support.
Encode states so scientifically relevant structure can be preserved.
Make the relations that may drive behaviour explicit and inspectable.
Update structure from evidence while tracking stability and uncertainty.
Return an explanation, control, or design implication to evidence.
Scope
These choices determine which distinctions a system can preserve, which relationships it can express, and how it can use limited evidence.
Organise molecular, sequence, material, or multimodal states around distinctions that matter to the scientific question.
Design model architectures and information flow to preserve scientific distinctions and express relevant relationships.
Use labels, constraints, perturbations, and learning objectives that reflect the intended scientific question.
Study how representations and interactions stabilise, transfer, fail, and change as new evidence arrives.
Study how models, data, simulation, experiments, tools, and scientific judgment work together, and what can be tested, compared, and reused.
Assess progress toward scientific understanding and reliable design.
Research contributions
We develop representations, interaction models and evaluation methods that connect learning architecture to scientific questions.
Representative testbeds
Testbeds are used to challenge transfer, identifiability, and evidence requirements. They do not redefine the Pillar as an applications category.
Research horizon
These are planning windows. Expansion will follow research progress and validation.
We develop and test AI methods for mechanism understanding, discovery, and design through scientific questions and feedback from experiments and simulation.
We aim to develop reusable representation and model methods, scientific system architectures, research infrastructure, evidence-acquisition strategies, and evaluation systems that support mechanism understanding and reliable design.
We aim to extract mathematical, computational, and system-design principles and test how these methods and capabilities transfer and combine across broader scientific fields.
Scientific principles
Use strict splits, matched controls, and null tests that challenge easier alternative explanations.
Treat sensitivity, uncertainty, negative evidence, and boundary conditions as part of the scientific result.
Interpretation should imply a perturbation, control, observation, or design decision that could be wrong.
Connect conclusions to the data, methods and analyses that support them.
Continue through the system
