Closed-Loop Scientific Discovery

• Closed-loop scientific discovery is an adaptive framework that iteratively integrates hypothesis generation, experimental design, data acquisition, and model updates under explicit feedback control.
• Its methodology emphasizes iterativity, adaptivity, and resource-aware experiment selection to maximize sample efficiency across domains such as materials discovery and drug design.
• Key applications leverage Bayesian optimization, reinforcement learning, and dynamic knowledge updating to drive autonomous and efficient scientific exploration.

Closed-loop scientific discovery refers to the iterative, adaptive, and resource-constrained integration of hypothesis generation, experimental (or computational) design, observation, and belief/model update, all under explicit feedback between computational reasoning and empirical execution. The paradigm stands in contrast to “open-loop” workflows, which are static or purely predictive, and operationalizes the scientific method in closed feedback cycles designed for maximally efficient and autonomous exploration of high-dimensional hypothesis or candidate spaces (Malik et al., 28 Jan 2026, Zenil et al., 2023, Duraisamy, 26 Jun 2025, Kramer et al., 2023). Modern closed-loop systems combine algorithmic hypothesis generators, model- or policy-driven experiment selection, real or simulated data acquisition, and critical loop closure in the form of adaptive reasoning, retraining, or formal hypothesis curation. These systems now span domains from materials discovery and drug design to symbolic law recognition, causal inference, and fundamental science.

1. Formal Structure and Key Principles

Closed-loop scientific discovery is built on three defining characteristics (Malik et al., 28 Jan 2026, Duraisamy, 26 Jun 2025, Kramer et al., 2023):

1. Iterativity: At each iteration, all data collected so far are used to inform the next experimental or computational action; none of the pipeline is precomputed.
2. Adaptivity: Action, hypothesis refinement, and resource allocation are reactive to empirical feedback, including both successes and failures.
3. Resource Constraints: Experiment/simulation budgets are finite, typically due to cost, time, or throughput bottlenecks, necessitating sample-efficient and judicious action selection.

This loop is mathematically formalized by policies $\pi$ mapping the history $\mathcal{H}_t$ or belief state to experimental action $a_t$:

$$a_t \leftarrow \pi(\mathcal{H}_t),\quad \mathcal{H}_t = \{ (a_i, o_i) \}_{i=1}^t$$

where $o_i$ are the observed outcomes. Bayesian or information-theoretic objectives predominate:

$$a_t = \underset{a\in\mathcal{A}}{\arg\max}~ \mathbb{E}_{o\sim P(\cdot|a)} \Big[ U(\pi, \mathcal{H}_t, a, o) \Big]$$

with $U$ typically reflecting expected information gain, improvement, or uncertainty reduction (Kramer et al., 2023, Kusne et al., 2020, Duraisamy, 26 Jun 2025, Jain et al., 2023). The cycle closes through posterior/model update and adaptive generation of new hypotheses.

2. Common Architectures and Computational Building Blocks

The architecture of a closed-loop discovery system typically decomposes into interacting modules (Zenil et al., 2023, Duraisamy, 26 Jun 2025, Malik et al., 28 Jan 2026, Mei et al., 20 May 2026):

Stage	Function	Example Implementation
Hypothesis Generator	Propose models / candidates	LLMs, GFlowNets, generative models
Experiment/Simulation Selector	Choose next actions under uncertainty/resource limits	Bayesian optimization, RL, planners
Data Acquisition/Execution	Enact measurement or simulation	Robotics, remote orchestration
Outcome Analysis	Score, filter, and evaluate results	Surrogates, uncertainty quantification
Model Update/Refinement	Update beliefs, models, or knowledge base	Bayesian updates, knowledge graphs
Loop Controller	Orchestrate full cycle, interface with user/governance	Agents, human-in-the-loop, governance

A key feature is composability: modular agents for generation (e.g., crystal structure generators, LLMs), filtering and selection (e.g., chemical validity, property screens), and planning (e.g., LLM agentic orchestration or learned acquisition policies) (Malik et al., 28 Jan 2026, Weng et al., 30 Sep 2025, Wang et al., 18 May 2026). Persistent “memories” or knowledge graphs allow continual learning from the experimental sequence (Duraisamy, 26 Jun 2025, Zenil et al., 2023, Mei et al., 20 May 2026).

3. Mathematical Formalization and Optimization Objectives

Closed-loop objectives are sharply formulated for different domains:

Materials Discovery (as in MADE)

• Given a chemical search space $S$ and oracle $O:S\to\mathbb{R}$, the agent seeks to maximize the number of unique, stable, and novel materials found under a budget $B$.
• Stability—using convex hull construction:

$\mathcal{H}_t$0

with discovery defined by

$\mathcal{H}_t$1

• Evaluation metrics include mSUN, AUDC, Acceleration/Enhancement Factors (Malik et al., 28 Jan 2026).

Information-Theoretic or Bayesian Active Learning

• Acquisition function given by expected improvement (EI), upper confidence bound (UCB), or information gain:

$\mathcal{H}_t$2

$\mathcal{H}_t$3

Symbolic and Mechanistic Discovery

• Model selection/posterior over models $\mathcal{H}_t$4 or hypotheses $\mathcal{H}_t$5:

$\mathcal{H}_t$6

$\mathcal{H}_t$7

4. Exemplary Systems and Empirical Results

MADE: Modular Benchmark for Materials

• Agentic discovery pipelines combining structure generators (random, Chemeleon), surrogate models (MLIP), LLM-based planners, and orchestration policies outperform hand-designed pipelines and random baselines in stable compound discovery. Chemeleon+MLIP achieves $\mathcal{H}_t$8 (discovery acceleration) (Malik et al., 28 Jan 2026).

DrSR: Dual-Reasoning Symbolic Regression

• Closed-loop LLMs combining explicit data-driven structural analysis and reflective learning yield >99% accuracy and two orders of magnitude lower NMSE versus non-closed-loop LLMs in symbolic equation recovery (Wang et al., 4 Jun 2025).

DeepScientist: Autonomous Discovery as Bayesian Optimization

• Multi-stage (hypothesize, verify, analyze) closed loops, each filtering and promoting only high-value findings, recorded >183% improvement over human SOTA in select AI tasks, albeit with <3% overall yield per candidate and heavy computational cost (Weng et al., 30 Sep 2025).

LLM-AutoSciLab and LLM-ACES: Symbolic & Graph Recovery

• Active, hypothesis-conditioned experiment selection achieves 2–5× fewer queries for mechanism or network recovery versus static baselines; symbolic accuracy up to 67.6% on NewtonBench and 31.1% graph exact recovery on gene network tasks (Kabra et al., 21 May 2026, Abhyankar et al., 23 Jun 2026).

AIMBio-Mat: Multiobjective, Uncertainty-Aware Biomedical Materials

• AI-native platform leveraging knowledge graphs, human review, Bayesian multiobjective optimization, and full data/decision audit trails. Operationalizes risk-tiered governance and can close the feedback loop across data ingestion, model training, experiment, and review (Mei et al., 20 May 2026).

5. Metrics and Performance Evaluation

Closed-loop scientific discovery evaluations move beyond predictive accuracy to include:

• Discovery-Centric Metrics: Number of unique, stable, and novel candidates (e.g., mSUN), area under the discovery curve (AUDC), acceleration/enhancement factors (Malik et al., 28 Jan 2026).
• Generalization and Transferability: Held-out test set validation (discovery vs. certification separation); audit of non-transfer signatures due to selection variance or distribution shift (Ning et al., 22 Jun 2026).
• Efficiency and Resource Use: Sample efficiency (queries to solution), experiment time, closed-loop efficiency relative to human or random baselines (Malik et al., 28 Jan 2026, Zenil et al., 2023).
• Human-AI Orchestration: Effective incorporation of human-in-the-loop decision gates, audit trails, and governance for high-stakes experimental recommendations (Mei et al., 20 May 2026).
• Explainability and Scientific Insight: Rule derivation, causal mechanism encoding, and convergence to interpretable models (Ji et al., 23 Sep 2025, Zenil et al., 2023, Jagadish et al., 24 Jun 2026).

6. Scalability, Generalization, and Limitations

Contemporary systems scale from tens (protein/ligand discovery) to millions (chemical/materials) of candidates, and support multi-objective, multi-fidelity, and multi-domain loops with modular agentic architectures (Jain et al., 2023, Malik et al., 28 Jan 2026, Mei et al., 20 May 2026). Scaling challenges include:

• High computational costs for fine-grained or high-dimensional Bayesian optimization (e.g., 20K GPU-hours in DeepScientist (Weng et al., 30 Sep 2025)).
• Brittleness to distribution shift or noisy validation signals; need for clear separation between loop optimization and hypothesis certification (Ning et al., 22 Jun 2026).
• Human judgment remains a required architectural component in ambiguous, ethically sensitive, or regulatory-adjacent applications (Duraisamy, 26 Jun 2025, Mei et al., 20 May 2026).
• Scientific AI-readiness and semantic data integration are necessary for high-fidelity, autonomous execution in new instrumental or lab settings (Rao et al., 9 Feb 2026).

7. Outlook and Future Directions

Open grand challenges for closed-loop scientific discovery include (Kramer et al., 2023, Duraisamy, 26 Jun 2025, Zenil et al., 2023):

• Integration of symbolic regression and mechanistic model discovery inside full-stack self-driving experimental labs.
• Unified frameworks merging deep learning, Bayesian planning, uncertainty quantification, and symbolic/neuro-symbolic reasoning.
• Democratization and open-source “AI-native” science platforms with standardized FAIR data and rigorous audit/governance.
• Advancement from domain-tuned, single-objective loops to cross-domain, Level 4–5 fully autonomous scientific reasoning—enabling AI scientists to propose new questions, invent new devices, and synthesize human-level breakthroughs.

Closed-loop scientific discovery thus defines the operational core of “automated science,” blending algorithmic and human agents in adaptive, resource-efficient, and epistemically rigorous cycles—now validated across a spectrum of real-world applications from molecular property prediction to cognitive theory construction (Kramer et al., 2023, Jagadish et al., 24 Jun 2026, Zeng et al., 25 Jun 2026, Malik et al., 28 Jan 2026).