Source-linked AI summary
On scientific understanding with artificial intelligence
Mario Krenn, Robert Pollice, Si Yue Guo, Matteo Aldeghi, Alba Cervera-Lierta, Pascal Friederich, Gabriel dos Passos Gomes, Florian Häse, Adrian Jinich, AkshatKumar Nigam, Zhenpeng Yao, Alán Aspuru-Guzik
TL;DR
The paper asks how artificial systems can contribute to or achieve scientific understanding beyond prediction and rediscovery. It combines philosophy of science, literature review, and scientists’ accounts to define three dimensions of android-assisted understanding, illustrated through computational simulations and machine-generated insights, while identifying requirements for future agents of understanding.
Problem
Artificial systems can rediscover laws or predict outcomes, but it remains unclear whether they provide scientific understanding beyond what their human creators already seek.
Method
The paper combines a literature review with surveys of dozens of scientists and uses these accounts to define three dimensions of android-assisted scientific understanding.
Results
The paper characterizes artificial systems as computational microscopes, resources of inspiration, and a not-yet-existent agent of understanding, supported by examples from molecular dynamics and neural-network analysis.
Takeaways & Limitations
Future progress requires collaboration among natural scientists, computer scientists, and philosophers to develop androids that directly contribute to scientific understanding.
Takeaways & Limitations
Current systems such as AlphaFold remain black-box oracles that do not directly provide scientific understanding.
Abstract
from arXiv · showhide
Imagine an oracle that correctly predicts the outcome of every particle physics experiment, the products of every chemical reaction, or the function of every protein. Such an oracle would revolutionize science and technology as we know them. However, as scientists, we would not be satisfied with the oracle itself. We want more. We want to comprehend how the oracle conceived these predictions. This feat, denoted as scientific understanding, has frequently been recognized as the essential aim of science. Now, the ever-growing power of computers and artificial intelligence poses one ultimate question: How can advanced artificial systems contribute to scientific understanding or achieve it autonomously? We are convinced that this is not a mere technical question but lies at the core of science. Therefore, here we set out to answer where we are and where we can go from here. We first seek advice from the philosophy of science to understand scientific understanding. Then we review the current state of the art, both from literature and by collecting dozens of anecdotes from scientists about how they acquired new conceptual understanding with the help of computers. Those combined insights help us to define three dimensions of android-assisted scientific understanding: The android as a I) computational microscope, II) resource of inspiration and the ultimate, not yet existent III) agent of understanding. For each dimension, we explain new avenues to push beyond the status quo and unleash the full power of artificial intelligence's contribution to the central aim of science. We hope our perspective inspires and focuses research towards androids that get new scientific understanding and ultimately bring us closer to true artificial scientists.
I. INTRODUCTION
The paper asks how artificial systems can contribute to scientific understanding rather than merely produce predictions or discoveries. It combines literature, philosophy of science, and scientists’ accounts to define three dimensions of android-assisted understanding and propose future directions.
- Artificial intelligence is proposed as a potential new tool for scientists, but its ability to contribute fundamentally to scientific understanding remains questioned.
- The authors combine a literature review with narratives from dozens of scientists working across biology, chemistry, physics, artificial intelligence, and advanced computational methods.
- Using a philosophy-of-science framework, the paper introduces three dimensions of android-assisted scientific understanding.
- In the computational-microscope dimension, androids reveal otherwise difficult-to-probe properties, which humans then elevate into scientific understanding.
- In the resource-of-inspiration dimension, androids generate concepts and ideas that human scientists subsequently understand and generalize.
- In the agent-of-understanding dimension, the machine itself would gain understanding, although the authors report no evidence that such agents yet exist.
- The first two dimensions help humans gain understanding, whereas the last requires the machine to gain understanding itself.
II. SCIENTIFIC UNDERSTANDING
The paper adopts a contextual and pragmatic theory in which understanding requires intelligible theories whose characteristic consequences scientists can recognize without exact calculations. This framework makes understanding experimentally assessable through scientific outcomes and consequences.
- De Regt and Dieks describe scientific understanding as contextual and pragmatic, treating visualization and unification as tools for understanding.
- A phenomenon is understood when an intelligible theory exists and scientists can recognize its qualitatively characteristic consequences without exact calculations.
- The framework separates understanding into a criterion for phenomena and a criterion for the intelligibility of theories.
- The authors use this theory to evaluate understanding through scientists’ successful outcomes and consequences rather than by inspecting their methodology.
III. WHAT IS NEXT?
The paper argues that rediscovering known scientific laws or concepts is not enough to establish new scientific understanding. It therefore calls for research that explicitly targets understanding beyond rediscovery.
- Rediscovery tasks do not guarantee new scientific understanding because human choices about targets, representations, code, and data analysis may introduce conscious or unconscious biases.
- The authors therefore advocate moving beyond rediscovery tasks and focusing explicitly on how artificial systems can produce new scientific understanding.
B. Beyond Discovery
Prediction and discovery can produce major scientific or technological advances without directly providing scientific understanding. The paper therefore argues that AI research should pursue concepts humans can apply beyond complete computation.
- Game-changing discoveries may not qualify as scientific understanding if scientists cannot use their underlying principles in other contexts.
- AlphaFold is described as a powerful protein-folding oracle whose black-box operation does not directly provide new scientific understanding.
- The paper’s ultimate goal is to obtain ideas or concepts from androids that can be applied in different situations without complete computations.
- The article aims to clarify this goal, assess previous approaches, and identify ways to advance AI’s role in natural-science research.
IV. THREE DIMENSIONS OF COMPUTER-ASSISTED UNDERSTANDING
The paper classifies computer-assisted scientific understanding into three independent, non-exclusive dimensions. The first two support human understanding, while the third envisions machines generalizing and transferring concepts themselves.
- The classification uses scientific literature, scientists’ anecdotes, and philosophy of science to map unexplored contributions to scientific understanding.
- I) Computational microscope: The computational microscope provides information not yet attainable through experiments.
- II) Resource of inspiration: The resource of inspiration, or artificial muse, expands human imagination and creativity.
- In the first two dimensions, human scientists remain essential for developing computational insights and inspiration into full understanding.
- III) Agent of understanding: The agent of understanding would replace humans in generalizing observations and transferring scientific concepts to new phenomena.
- The three classes are intended as guides to future possibilities rather than dogmatic categories.
A. Computational microscope for scientific understanding
Computational microscopes simulate biological, chemical, or physical processes beyond experimental perception. Their outputs can yield generalizable biological concepts that support understanding without rerunning complete simulations.
- Computational microscopes investigate objects or processes that cannot be visualized or probed otherwise, including phenomena beyond experimental length and time scales.
- Computer-generated data must be generalized to other contexts without complete computation to contribute to scientific understanding.
- Molecular dynamics simulations of SARS-CoV-2 revealed different biological functions for open and closed spike-protein conformations, changing views of glycans.
- Glycoblocks provide patterns for understanding biomolecular sequence–structure–property relationships and designing synthetic structures without simulating entire systems.
The next computational microscope
The next computational microscope should combine more complex simulations with more interpretable representations of their data. The paper also proposes multisensory interfaces to help scientists detect structure and patterns.
- Future computational microscopes should analyze more complex physical systems and represent information more interpretably.
- More Complex Systems: Figure 2 identifies GPUs, TPUs, OPUs, and ultimately quantum computers as possible paradigms for larger and more complex computations.
- More Complex Systems: Larger systems, longer time scales, and more modeled interactions could increase simulation applicability, using algorithmic or hardware improvements.
- Full spectrum of senses: Three-dimensional virtual or augmented-reality environments could help scientists extract more from computer-generated data.
- Full spectrum of senses: Auditory, tactile, olfactory, and gustatory channels are proposed as additional ways to experience structure in scientific data.
B. Resource of inspiration for scientific understanding
Computers can surprise human scientists and serve as artificial muses, provoking ideas that may contribute to scientific understanding.
- Computer algorithms can systematically provoke surprising and creative ideas, potentially accelerating scientific and technological progress.Turing recognized that computers could surprise their creators, and studies of artificial life and evolution document such effects.
- Researchers’ accounts show that computer algorithms can surprise their human creators and produce behavior described as creative.
- The proposed computational muse would generate surprises and help humans use algorithmic outputs as sources of new scientific ideas.The paper distinguishes provoking surprising algorithmic behavior from humans lifting resulting insights into scientific understanding.
The future resource of inspiration
The future resource of inspiration is an android that systematically finds surprising data, literature, model behavior, agent behavior, or interpretable solutions to inspire human scientific concepts.
- Identifying surprises in data: Future systems could identify exceptional data points or unexpected regularities from experiments and simulations to inspire new scientific ideas.Closed-loop systems could steer computational or experimental exploration toward unexpected regions, requiring access to complex laboratory automation for experimental sources.
- Identifying surprises in data: Current examples show humans conceptualizing unexpected crystal structures and unusually large quantum entanglement discovered through computational searches.The examples yielded the concepts of spontaneous ionization and a new principle of entanglement generation.
- Identifying surprises in data: Autonomous anomaly detection could identify new physics signatures or hidden regularities that human scientists can then conceptualize and understand.Examples include applications to Large Hadron Collider data and mathematical discoveries involving previously unconnected knot-theory invariants.
- Identifying surprises in data: No reported case yet shows an A.I. uncovering previously overlooked scientific patterns or irregularities that lead to new conceptual understanding.
- Identifying surprises in the scientific literature: Computers could mine the expanding scientific literature to identify exceptional phenomena, interdisciplinary connections, islands, and unexplored regions.Semantic knowledge networks represent scientific concepts as nodes and their relations as edges, enabling systematic exploration of literature.
- Inspecting models and probing artificial agents: Inspecting models and probing artificial agents could reveal learned variables, design principles, thermodynamic properties, or unexpected solutions.Disentangled neural-network variables have been used to expose internal representations, while intrinsic curiosity, creativity, and surprise can drive exploration.
- New concepts from interpretable solutions: Interpretable solutions such as mathematical formulas or graphs can expose new concepts and transferable reasons why computationally discovered solutions work.Examples include symbolic models from physical data, rediscovery of Newton’s gravitation law, and graph-based designs for unknown quantum systems.
C. Agent of Understanding
The paper describes a future agent of understanding that autonomously acquires scientific understanding, recognizes theory-based consequences, and transfers that understanding to human experts.
- No surveyed respondent or published study has yet described an algorithm that autonomously acquires new scientific understanding.The paper therefore lists requirements, proposes detection tests, and speculates about such agents’ possible forms.
- Scientific novelty is context-dependent, so an understanding agent must evaluate whether an insight is new within at least a specific scientific domain.
- Naive neural-network prediction is insufficient because scientific understanding requires theories that reveal qualitatively characteristic consequences.Applying a learned conceptual core elsewhere does not by itself establish understanding without an explanation in a scientific theory.
- Condition I requires an android to recognize qualitatively characteristic consequences of a theory without exact computation and use them in a new context.
- Condition II requires an android to transfer its understanding to a human expert.
- The Scientific Understanding Test compares explanations from a human or android teacher through independent evaluation of non-trivial explanations across contexts.Understanding is attributed when a referee cannot distinguish the teacher’s explanations from those of a human scientist.
- The envisioned systems would combine advanced human-computer interaction with computational microscopes so humans can understand and interrogate machine-devised concepts.Natural-language scientific queries are proposed as one way for scientists to probe an algorithm.
V. CONCLUSION
The paper argues that artificial intelligence can contribute directly to acquiring scientific understanding, but future progress will require multidisciplinary collaboration.
- Advanced computational methods and artificial intelligence can contribute directly to acquiring new scientific understanding, a main aim of science.The authors call for collaboration between natural scientists and computer scientists to advance this use of androids.