SE-Fab: Self-Evolving Foundry for the Embodied AI Scientist
Special Issue Vol. 4 · AI4S Infrastructure · Scale Agentic Discovery
SE-Fab is an AI4S infrastructure concept for an autonomous enzyme-evolution laboratory: it turns research intent into schedulable, executable, traceable work, then uses real experimental feedback to generate a better next round.
From EvoPlay to the embodied robot scientist
As AI for Science enters the Agent era, scientific competition is shifting. The question is no longer who owns more tools, but who can place tools, data, equipment, and experimental feedback inside one continuously evolving loop.
SE-Fab—short for Self-Evolving Foundry—is proposed as the infrastructure for that loop: a self-evolving dry–wet laboratory for AI4S.
The loop: from research intent to the next experiment
SE-Fab connects the full cycle:
- Define a research objective.
- Design candidate experiments.
- Schedule the required resources.
- Execute the experiment.
- Record the process and its context.
- Learn from the feedback.
- Generate a better plan for the next round.
The laboratory does not stop at answering a question. It turns research intent into a task chain that is schedulable, executable, and traceable.
Why automation alone is not enough
Automation solves the problem of making an experiment run. The deeper R&D bottleneck often appears between experiments: deciding what to try next, learning from failures, and turning results into a better search strategy.
SE-Fab combines multi-agent collaboration, self-play, active-learning iteration, Agent-ready hardware, end-to-end traceability, safety and compliance, intelligent scheduling, and flexible dispatch.
Traditional approaches are constrained by limited throughput, rigid workflows, data silos, and difficulty applying a huge candidate space. A self-evolving lab learns from every round—not only successful experience, but also failure cases and the boundaries they reveal.
How SE-Fab works
1. Turn research intent into executable tasks
SE-Fab translates a research objective into a chain of tasks that can be dispatched, executed, and tracked. It can propose multiple experimentally testable protocols rather than returning only a conversational answer.
The image highlights four reported improvements for an enzyme self-evolution process compared with a traditional method:
- 27% fewer invalid experiments
- 30% higher equipment utilization
- 60% shorter iteration time
- 58% lower audit cost
2. Schedule intelligently so AI-designed plans reach the physical lab
Before dispatch, SE-Fab checks resource availability. An intelligent scheduling Agent then places each task into a time window that the real production line can run.
3. Make every device an Agent-callable experiment node
In an Agent-ready production line, each device exposes a standardized interface. AI Agents can call, schedule, and orchestrate those devices directly. Instruments stop being isolated islands and become experiment nodes that can be invoked whenever the workflow needs them.
The result is a laboratory that is traceable, auditable, and able to keep evolving.
4. Keep a trustworthy execution record
SE-Fab is not a black box that “automatically runs experiments.” It records the path from model suggestion to expert confirmation, from instrument action to data write-back, including:
- Exception records.
- Material batches and plate positions.
- Equipment status.
- Expert confirmations.
- Algorithm-optimization history.
- Signature and audit records.
Data returning to the system is not the end of archiving. It is the starting point for the next evolution cycle. Successful data becomes experience; failed data becomes a boundary. Every experiment returns with its context and quality labels, ready to become an asset for the next design.
From enzyme engineering to more AI4S laboratories
The same infrastructure can extend beyond enzyme engineering to synthetic biology, chemicals, materials, testing, and other R&D environments.
The execution chain remains consistent:
AI design → experiment orchestration → automatic execution → self-evolving iteration → data return → resource scheduling
The long-term goal is an embodied AI scientist: a system that can design, run, observe, learn, and improve inside a real laboratory.
Build your embodied AI scientist
SE-Fab’s proposition is concise: make the real laboratory part of the AI4S infrastructure, and make every experiment a step toward a more capable next experiment.
Build Your Embodied AI Scientist.
Reference image note: one source image cites a Nature Machine Intelligence article dated August 8, 2023, DOI 10.1038/s42256-023-00691-9.
Source: Yongsheng Intelligence, “【AI智能体|自主酶进化实验室】SE-Fab，Self-Evolving Foundry”