Self-Driving Discovery & Design

• Self-driving discovery and design is the integration of automated experiments, AI-guided decision making, and closed-loop feedback that accelerates research across drug discovery, materials science, and more.
• The approach leverages layered architectures that combine robotics, digital twins, and heuristic scheduling to optimize experimental protocols via Bayesian and generative models.
• These systems achieve dramatic efficiency gains with reproducible, model-based experimentation and robust safety mechanisms, while maintaining critical human-machine collaboration.

Self-driving discovery and design refers to the cyber-physical integration of automated experimentation, AI-driven decision-making, and closed-loop data-model feedback to autonomously explore, optimize, and generate new scientific knowledge or functional products. Originating in domains such as drug discovery, materials science, catalysis, and mechanistic engineering, self-driving laboratories (SDLs) now embody a unified paradigm for accelerating iterative scientific workflows, transcending traditional trial-and-error with model-informed, autonomous execution. Modern SDLs tightly coordinate robotics, instrumentation, workflow scheduling, generative AI models, Bayesian optimization, and FAIR data management to minimize experimental budgets, maximize throughput, and reliably produce or interpret discoveries without continuous direct human supervision.

1. Architectural Foundations of Self-Driving Laboratories

Self-driving discovery platforms universally center on a layered architecture that orchestrates laboratory hardware, workflow automation, AI/ML inference, and data management in real time. For drug discovery, the Artificial system exemplifies this hierarchy, exposing a core orchestrator and scheduler interfacing via Web Apps, a Lab API (GraphQL/gRPC/REST), and adapter modules to instruments, robots, and AI engines. This enables seamless interaction between workflow planners (Python/GUI), resource schedulers (heuristic and constraint-based), and data recorders, all underpinned by integration with LIMS, ELN, and cloud/on-premises data lakes. Experimental tasks—pipetting, incubation, analytic/AI inference—are modeled as tasks in a DAG with resource and capacity constraints, supporting complex dependencies, batching, and manual intervention via real-time digital twins (Fehlis et al., 1 Apr 2025).

Cross-domain SDLs, including the RAISE wettability lab, A-Lab GPSS for air-sensitive synthesis, and MCP-based NIMO Controller, further standardize device abstraction (JSON-Schema/MCP), workflow representation (visual programming or protocol DSLs), and orchestration of human-in-the-loop versus fully autonomous execution (Nazeri et al., 8 Oct 2025, Fei et al., 13 Apr 2026, Yoshikawa et al., 13 May 2026). At scale, distributed SDLs leverage digital twins and FAIR repositories (such as nanoHUB ResultsDB) for collaborative, geographically dispersed optimization (Deucher et al., 24 Jun 2025).

2. Workflow Automation, Scheduling, and Protocol Formalization

Automation of experimental design and execution leverages formal scheduling and protocol translation. The Artificial system schedules tasks in resource- and precedence-constrained DAGs, optimizing the makespan:

$$\min_{s_i}\,C_{\max} = \min_{s_i}\,\max_{i\in T} (s_i + p_i)$$

subject to precedence ($s_j \ge s_i + p_i$), resource capacity ($$\sum_{i:\,r_k\in \rho(i)} \mathbf{1}_{[s_i,\,s_i+p_i)}(t)\le C_k$$), and batching ($$|\{i \in \mathcal B: \text{batch}(\mathcal B) = b\}| \le B_{\max}$$) constraints, with heuristic and mathematical programming solvers yielding near-optimal schedules within seconds (Fehlis et al., 1 Apr 2025).

For protocol knowledge exchange, frameworks such as Protocol Dependence Graphs (PDG) automate translation from natural language to machine-executable formats. PDGs explicitly encode syntax (control flow), semantics (data/reagent flow), and execution constraints (resource, spatial-temporal, and safety) to enable full verification and simulation before downstream automation. A staged workflow incrementally builds these graphs—parsing syntax, interpreting semantics, and simulating execution under capacity and safety limits—with quantitative evaluation showing parity or slight inferiority to human-expert translation but vastly increased throughput (Shi et al., 2024). Hierarchical encapsulation using domain-specific languages at multiple abstraction levels (instance action, generalized operation, product flow model) supports LLM-guided design, modification, and adjustment of new protocols, systematically enforcing intra- and inter-step soundness (Shi et al., 4 Apr 2025).

3. Closed-Loop, AI-Integrated Experimental Design

Self-driving discovery is predicated on closed AI/ML-driven loops, fusing model-driven hypothesis generation, experimental execution, high-throughput data acquisition, and continuous model update. High-level workflow patterns include:

1. Design: Define objective criteria (e.g., protein target, material property) and initialize experimental or candidate libraries.
2. Run: Execute experiments in batches according to scheduler-determined resource allocation and protocol steps, leveraging sensing, imaging, computation, and analytics.
3. Optimize/Learn: Integrate outputs into surrogate or generative models; employ acquisition functions to select subsequent queries for exploitation and exploration.
4. Update: Retrain (or incrementally update) surrogates, acquisition policies, and candidate pools, closing the feedback loop.

In drug discovery, AI models (e.g., NVIDIA BioNeMo) provide molecular property inference (binding affinity ΔG), driving Bayesian optimization with surrogate models ($\hat f(x|D)$), acquisition functions ($\alpha(x)=EI(x), UCB(x)$), and generative filtering, with rounds of wet-lab and in silico validation (Fehlis et al., 1 Apr 2025). Adaptive microscopy platforms employ dual-novelty deep kernel learning (DN-DKL) for balancing spectral/structural novelty and model uncertainty, and dual-VAE architectures for multimodal latent space embedding, efficiently mapping structure-property relationships and directing acquisition to maximize informative content (Gong et al., 17 Mar 2026).

Human expertise remains integral in novel or hard-to-scalarize regimes. Deep-kernel pairwise learning (DKPL) incorporates expert pairwise judgments (not explicit scalar scores) to learn latent utility functions over high-dimensional data (e.g., nanoscale images), thus enabling expert-guided experiments where key phenomena are non-numeric or multi-objective (Bulanadi et al., 20 May 2026).

4. Model-Based Optimization, Active Learning, and Experiment Planning

AI-accelerated self-driving discovery hinges on statistical and algorithmic frameworks for efficient resource allocation and search. The dominant regime is Gaussian process (GP)-based Bayesian optimization (BO), applied pervasively in chemical, materials, and process engineering SDLs (Adesiji et al., 8 Aug 2025, Liang et al., 24 Feb 2026, Nazeri et al., 8 Oct 2025, Advincula et al., 26 Feb 2026, Xu et al., 2 Sep 2025). A generic GP surrogate is

$f(x)\sim \mathcal{GP}(m(x), k(x, x'))$

with closed-form expressions for the posterior mean/variance, and acquisition strategies such as Expected Improvement (EI), Upper Confidence Bound (UCB), or Thompson sampling. Multi-objective extensions, constraint-aware optimization (e.g., using penalization or Pareto dominance), and batch querying are implemented in high-throughput campaigns (Fehlis et al., 1 Apr 2025, Nazeri et al., 8 Oct 2025, Advincula et al., 26 Feb 2026).

In case studies, SDLs routinely achieve quantitative acceleration metrics. For example, the thin-film synthesis SDL reduced the number of required experiments by over 30× (27 BO-guided cycles versus ≈1000 grid points) (Liang et al., 24 Feb 2026). RAISE achieved up to a 3× improvement in sample efficiency for multi-objective formulation via BO compared to single-objective runs (Nazeri et al., 8 Oct 2025). Review benchmarks report an overall median acceleration factor (AF) of 6, with peak enhancement factors (EF) at around 10–20 experiments per problem dimension, consistently demonstrating a “blessing of dimensionality” for BO over random search (Adesiji et al., 8 Aug 2025).

Customizations appear according to property domain: in catalysis, feature engineering draws on mechanistic descriptors (surface energies, d-band centers, sequence embeddings), while reinforcement learning handles multi-step syntheses (Advincula et al., 26 Feb 2026). Generative approaches leveraging VAEs and neural network equation learners (nn-EQL) extract interpretable structure–property laws directly from experimental data, streamlining scientific inference (Desai et al., 2024, Desai et al., 2024).

5. Performance Evaluation and Quantitative Metrics

The quantification of SDL efficiency involves throughput, acceleration, reproducibility, utilization, and discovery coverage metrics. Systems such as Artificial report:

• Throughput speedup:

$$S = \frac{T_{\text{manual}}}{T_{\text{auto}}} \approx 10\times – 100\times$$

• Reproducibility: Each run is immutably logged, yielding $R \ge 0.99$ re-execution fidelity.
• Resource utilization:

$$U = \frac{\sum_{i} p_{i}}{C_{\max}} \ge 0.85 \quad \text{(vs. 0.5 in uncoordinated labs)}$$

Autonomous materials discovery campaigns consistently reduce the number of experiments by an order of magnitude or more compared to grid or random search, with high hit rates and comprehensive coverage in parameter or compositional space (e.g., 72% pairwise metal combinations in 19-element solid-state synthesis) (Fei et al., 13 Apr 2026). In polymer optimization, self-driving platforms converged to LCST targets within 2–3 BO steps after initialization (Xu et al., 2 Sep 2025). Mechanical design SDLs explored >10¹¹ geometries, realized energy-absorbing structures at K* = 75.2%, and extracted transferable design heuristics (Snapp et al., 2023).

6. Safety, Robustness, and Human-Machine Integration

SDLs must address unique safety challenges distinct from conventional laboratory or digital-only AI systems. The Safe-SDL framework formalizes safety boundaries via:

• Operational Design Domains (ODDs): Constraint sets in laboratory state space, mathematically defining permissible regions (C, S) (Zhang et al., 13 Feb 2026).
• Control Barrier Functions (CBFs): Real-time enforcement that guarantees closed-loop invariance of safety envelopes, solved as constrained QPs at the hardware control layer.
• Transactional Protocols (CRUTD): Six-phase digital-to-physical transaction model (create, read/lock, simulate, test, execute, confirm), ensuring atomic, reversible, and verifiable execution, with automated aborts and logs on violation detection.

Empirical evaluations with platforms such as UniLabOS and Osprey demonstrate 100% intercept of dangerous AI-generated plans before hardware execution. The combination of digital simulation, formal verification, and edge-enforced CBFs closed the "syntax-to-safety gap" that foundation models alone fail to address (Zhang et al., 13 Feb 2026).

Human–machine synergy remains pivotal throughout: human experts oversee exception handling, hypothesis generation (via abductive/inductive LLMs), exploration/exploitation calibration, and validation of discovered knowledge. Human-in-the-loop verification is further critical in protocol translation, goal setting, high-risk operation gating, and cross-domain generalization (Shi et al., 2024, Fei et al., 13 Apr 2026, Bulanadi et al., 20 May 2026).

7. Impact, Collaborative Infrastructure, and Prospects

Self-driving discovery and design fundamentally accelerates the scientific discovery process, reduces cost, and drives reproducibility and robustness across chemistry, materials, biology, and engineering. The generalized adoption of FAIR data practices (findable, accessible, interoperable, reusable), cloud-native APIs (nanoHUB Sim2L), and open, plugin-based orchestrators (e.g., MCP/NIMO Controller) fosters distributed, collaborative SDL networks. These approaches standardize experiment representation, data exchange, and agent access, democratizing access to automation and enabling federated discovery at scale (Deucher et al., 24 Jun 2025, Yoshikawa et al., 13 May 2026).

SDLs are expanding in scope—integrating richer multimodal sensing, federated model sharing, and autonomous reasoning agents—while advancing interpretability (via symbolic regression) and safety engineering. Current research directions seek to embed deeper causal inference in protocol design, unify continuous and discrete domain representations, and enhance expert–AI collaboration in exploratory scientific contexts. Ongoing benchmarking and ablation studies continue to refine design choices, highlight domain-specific constraints, and calibrate expectations for SDL performance, robustness, and knowledge generation (Adesiji et al., 8 Aug 2025).

Self-driving discovery and design thus define an emergent mode of scientific practice—empirically grounded, statistically accelerated, and architecturally principled—where AI, automation, and model-informed feedback are co-engineered to deliver rapid, reproducible, and scalable innovation.