Bohrium+SciMaster: Agentic Science Ecosystem

• Bohrium+SciMaster is a unified infrastructure that integrates managed scientific assets, hierarchical models, and workflow orchestration for agentic science.
• The platform orchestrates long-horizon scientific workflows with multi-agent coordination and DAG-style traceability to significantly reduce cycle times.
• Deployments show dramatic speedup factors and improved auditability across diverse domains, ensuring rigorous integration of atomic-scale and large-scale data.

Bohrium+SciMaster designates an integrated infrastructure and orchestration framework for agentic science, wherein AI agents execute multi-step scientific workflows that interleave reasoning, tool use, and verification, thereby transforming isolated prototype efforts into an observable, reproducible, and evolvable Science-as-a-Service ecosystem. The stack comprises Bohrium—a managed, traceable hub for scientific assets; the Scientific Intelligence Substrate—a hierarchical ontology of models, knowledge, and workflow building blocks; and SciMaster—the orchestration engine that enables long-horizon workflow composition, execution, and audit. Representative deployments demonstrate orders-of-magnitude improvements in end-to-end scientific cycle time across diverse scientific domains, accelerated by execution-grounded feedback loops and large-scale capability sharing (Zhang et al., 23 Dec 2025). Bohrium+SciMaster also enables rigorous handling of atomic-scale data, facilitating integration of complex theoretical outputs such as those generated for Bohrium element spectroscopy (Lackenby et al., 2019).

1. Infrastructure Layer: Bohrium Hub

Bohrium functions as a managed hub transforming heterogeneous raw scientific assets—documents, datasets, simulation codes, compute clusters, and laboratory instruments—into agent-ready, traceable capabilities. The minimal contract for capabilities includes:

• Schema specification for inputs and outputs,
• Reproducible execution envelopes via version-controlled environments,
• Full run traceability supporting logging and provenance audit,
• Governed scheduling and resource accounting.

Key subsystems within Bohrium include:

• Science Navigator: Ingests multimodal evidence (papers, patents, figures, equations, chemical data) and binds workflow fragments to source documents, supporting semantic search and citation traversal.
• Lebesgue: Enables unified job submission and resource policy enforcement across compute backends; simulation engines (DFT, MD, FEM, graph nets) are exposed as callable services, each invocation generating a record $$r_i = (\text{tool\_id}, \text{inputs}, \text{outputs}, t_\text{start}, t_\text{end}, \text{status}, \text{cost})$$.
• UniLabOS: Virtualizes laboratory protocols and hardware, rendering wet-lab procedures schedulable, auditable, and transaction-safe.

Registries maintain versioned models/workflows/tools; platform-wide observability enables replay, debugging, and pipeline governance (Zhang et al., 23 Dec 2025).

2. Orchestration Layer: SciMaster Engine

SciMaster serves as the runtime environment for long-horizon scientific workflows. It enables explicit separation between reasoning, handled by tool-augmented models, and governed execution on Bohrium. Its principal features are:

• Task understanding and workflow construction: Reasoning models map natural-language or structured objectives to explicit pipelines of Reading, Computing, and Experiment steps, with declared dependencies and validation gates.
• Multi-agent coordination: Specialized agents (literature mining, simulation planning, experimental execution) interact under runtime guards (e.g., schema checks, parameter limits, quotas).
• Long-horizon state and memory: Versioned data artifacts enable branch comparison, rollback, and selection.
• DAG-style traceability: Each workflow is represented as $G=(V,E)$, logging invocations, artifacts, and validation results.

Validation gates enforce domain-specific correctness (e.g., mesh quality, patent-scope compliance), ensuring scientific rules are continuously checked during execution.

3. Scientific Intelligence Substrate

Bohrium+SciMaster is underpinned by a scientific intelligence substrate composed of hierarchical models, knowledge bases, and community assets:

• General-purpose foundation models (“Innovator”) for task interpretation and protocol/code generation.
• Domain-specific models (Uni-Mol, Uni-RNA, DPA) encoding modality priors for molecular, atomistic, or bioinformatic representations.
• Pipeline/application models (Uni-Parser, Uni-QSAR, Uni-AIMS): Optimized for execution stability and high-fidelity prediction in specialized scientific contexts.
• SciencePedia knowledge base: Contains structured concepts and long chains of thought (LCoT) with explicit provenance—enables conceptual trace-back and reuse.
• Community ecosystems (DeepModeling, etc.): Engine/solver codebases, workflows, and utilities contributed under standard conventions; examples include DeepMD-kit, ABACUS, dpdata, jax-fem, DP-GEN, APEX.

This substrate allows modular composition and systematic audit, with execution signals feeding back into continuous model/workflow improvement (Zhang et al., 23 Dec 2025).

4. Formal Metrics and Observability

End-to-end scientific workflow execution within Bohrium+SciMaster is measured via formally defined metrics:

• End-to-end cycle time: $$T(W) = \sum_{r_i \in W} (t_\text{end}^i - t_\text{start}^i)$$, e.g., $$T = T_\text{read} + T_\text{compute} + T_\text{experiment}$$.
• Speedup factor: $F = T_\text{baseline} / T_\text{agent}$; reduction $$R = (T_\text{baseline} - T_\text{agent}) / T_\text{baseline} \times 100\%$$.
• Trace completeness: $$C_\text{trace} = N_\text{recorded} / N_\text{expected} \in [0,1]$$.
• Execution-grounded signal rate: $R_\text{signal} = N_\text{signal} / \Delta t$; $\sigma = N_\text{signal} / N_\text{exec}$.

Deployment statistics indicate $N_\text{exec} = \mathcal{O}(10^6)$ workflow runs per month and millions of validation signals aggregated for agent improvement (Zhang et al., 23 Dec 2025).

5. Domain Workflows: Eleven Master Agents

Bohrium+SciMaster has been validated in production-scale deployments via eleven representative “Master Agent” workflows, each integrating SciMaster orchestration and Bohrium capabilities:

Agent	Scientific Domain	Sample Workflow Skeleton
AMTechMaster	Additive Manufacturing	CAD/NLP → geometry cleanup → meshing → FEM → stress analysis
FlowXMaster	CFD Simulation	Text/sketch → geometry reconstruction → mesh → solve → report
MatMaster	Materials Design	Lit/data mining → candidate gen. → DFT/ML → lab → data ingestion
ML-Master	ML Automation	Task parse → code gen → training → eval → refinement
OPT-Master	Optimization/OR	Problem desc. → model → solver → evaluation → refinement
PaSaMaster	Literature Search	Query → citation expansion → reasoning → result → provenance rep.
PDEMaster	Text-to-PDE Simulation	Text → weak form → mesh → FEM → validation → correction
PharmMaster	Patent Analysis	Patent parse → scaffold extraction → SAR synthesis → FTO assess
PhysMaster	Computational Physics	Assumption extraction → numerical scan → consistency check
SpecMaster	Structure Elucidation	Spectrum → feature extr. → candidate gen. → sim. → consistency
SurveyMaster	Survey Writing	Topic → retrieval → cluster → draft → citation validation

Each agent executes domain-specific tools, codes, and protocols through governed interfaces, with SciMaster enforcing validation gates and logging detailed execution traces.

6. Quantitative Impact and Ecosystem Advantages

Deployment of Bohrium+SciMaster has led to orders-of-magnitude improvements in scientific throughput:

• Literature search (PaSaMaster): $$T_\text{manual} \approx 120\text{ min} \to T_\text{agent} \approx 3\text{ min}$$, $F \approx 40\times$.
• Patent landscaping (PharmMaster): $$T_\text{manual} \approx 10\text{ days} \to T_\text{agent} < 1\text{ day}$$, $F \approx 10\times$.
• Survey writing (SurveyMaster): $$T_\text{manual} \approx 30\text{ days} \to T_\text{agent} \approx 4\text{ h}$$, $F \approx 180\times$.
• Materials design (MatMaster): Optimization cycle times reduced from months to days, $F \sim 10$–$30\times$, with up to 80% decrease in invalid-experiment rates.

Advantages over bespoke prototypes include systematic reuse of capabilities, side-by-side comparability, continuous improvement through aggregated execution signals, and platform-scale community participation. Uniform governance, validation, and trace logging anchor workflow observability and auditability in production settings (Zhang et al., 23 Dec 2025).

7. Example: Bohrium Element Spectroscopy Data Integration

Bohrium+SciMaster’s ecosystem facilitates the encoding, sharing, and exploitation of atomic-level data such as the relativistic electronic structure of Bohrium (Bh, Z=107):

• Level structure: Ground configuration [Rn] $5f^{14} 6d^5 7s^2$, $^6S_{5/2}$, $J=5/2$, $E=0\,\text{cm}^{-1}$, $g=1.78$. Numerous even- and odd-parity excited states are computed (e.g., $^4P_{3/2}$ at 13 062 cm⁻¹, $^2S_{1/2}^o$ at 12 792 cm⁻¹).
• Ionization potentials: IP sequence for neutral Bh I to Bh V: 8.03 eV, 19.0 eV, 26.2 eV, 36.8 eV. These values arise from configuration interaction (CI) and many-body perturbation theory (MBPT) calculations (Lackenby et al., 2019).
• Isotope shifts: Dominated by field-shift contributions; for Bh I: $$\Delta\nu(A_2, A_1) = a(A_2^{2\gamma/3} - A_1^{2\gamma/3})$$, $\Delta\nu = F \Delta \langle r^2 \rangle$. Example coefficients $a=-160$ cm⁻¹, $F=-33.0$ cm⁻¹ fm⁻² for transition $^6S_{5/2} \rightarrow ^6F_{5/2}^o$.
• Strongest E1 lines: Includes $^6F_{5/2}^o$ at 28 060 cm⁻¹ ($D_\text{E1}=1.51$ a.u., $A_\text{E1}=16.6 \times 10^6$\ s⁻¹).

Relativistic and spin–orbit effects are pronounced, with strong 6d–7s contraction/expansion, large $j$-splitting, and “jj-coupling” dominating over Hund’s rule. Bohrium+SciMaster supports protocolized ingestion and comparison of these theoretical data for cross-domain reasoning and experiment planning (Lackenby et al., 2019).

A plausible implication is that this infrastructure enables rapid, reproducible integration of atomic-scale calculations—such as those for Bohrium—into larger agentic workflows (e.g., nuclear radius extraction, isotopic shift modeling) within the broader Science-as-a-Service paradigm.

---

References

• Bohrium + SciMaster: Building the Infrastructure and Ecosystem for Agentic Science at Scale (Zhang et al., 23 Dec 2025)
• Theoretical study of electron structure of superheavy elements with an open 6d-shell, Sg, Bh, Hs and Mt (Lackenby et al., 2019)