The dry–wet closed loop is now complete and the full chain is connected. Oushisheng has delivered and deployed solutions across catalysis, pharmaceuticals, new materials and other application scenarios.
In recent years, the AI4S community has largely focused on developing and optimizing various large models. These models have continuously advanced artificial intelligence's upstream capabilities in chemistry, steadily improving the accuracy and speed of retrosynthetic planning, property prediction, and reaction condition screening, with a constant stream of achievements. This has seemingly created a prevailing consensus in the field: once model capabilities reach a sufficient level, the pace of scientific discovery can be substantially accelerated.
To progress from research conception to reproducible findings, an AI4S project requires six core steps organized across four tiers, encompassing the entire research pipeline. These steps include: scheduling computing resources at the compute layer, collecting and cleansing data at the data layer, training, inferring, and screening models at the model layer, and validating experiments at the experimental layer. Observational data from experiments is then fed back into the model system. Of these steps, the first four are considered "dry experiments," the fifth is the "wet experiment," and the sixth step determines the system's ability to self-iterate updates. This cyclical workflow is commonly known in academia as the AI4S dry-wet experiment closed loop.
In summary, current technological advancements primarily enhance the upstream "dry experiment" capabilities of AI systems. Several industry misconceptions persist:
· Believing computing scale is the main barrier is a false premise.
· Equating model performance directly with commercial competitiveness is misleading, as such advantages quickly diminish due to the rapid iteration of open-source ecosystems.
· Considering the sample size of public datasets and historical experimental data as core assets will only provide AI with an "illusion of successful pathways," failing to translate into real-world validation in practical chemical scenarios.
The industry's consistent neglect of the experimental layer often arises from the oversimplified assumption that authentic experimental data can be quickly and reliably acquired using experimental robots and automated instruments. In reality, the data generation rate from physical experiments is far too slow to meet the demands of continuous model training. Numerous critical bottlenecks must still be resolved for the effective implementation of the experimental layer.
Addressing the Bottlenecks: Understanding the Challenges of the Dry-Wet Closed Loop
What are the specific challenges in establishing the dry-wet closed loop? We will use a diagram to illustrate these issues.
Figure 1 | Dry-Wet Closed-Loop System Map
From an industry perspective, the first three tiers needed to operationalize the AI4S closed loop are largely available and unlikely to present significant barriers.
Computing resource scheduling has undergone a lightweight transformation. Hybrid schemes, combining local inference with on-demand cloud invocation, are increasingly replacing the traditional, capital-intensive deployment of full-time high-performance clusters. This approach optimizes the cost structure of computing investments. Competition at the compute layer has effectively ended; it has largely become a pure cost item, with comparisons focused on capital input rather than core capability barriers.
Agent-enabled efficiency gains are evident in data preparation. Repetitive tasks such as data cleansing, spectral interpretation, image recognition, parameter optimization, and data labeling can now be autonomously performed by intelligent Agents. This process transforms scattered, fragmented historical data into productive assets, accelerating their readiness for modeling.
The barriers to building model training infrastructure are continuously decreasing. As large models iterate rapidly, open-source base models, domain-specific databases, and supporting toolchains are maturing steadily. AI4S projects can now directly leverage mature model capabilities, significantly reducing the cost of building models from scratch. Consequently, the model layer is transitioning from "ground-up development" to training tasks focused on domain-specific adaptation and engineering tuning. Models themselves are no longer the primary advantage. They require vast amounts of real experimental data for training to adapt to diverse reaction systems and support closed-loop iteration. This involves automated model inference, screening of candidate molecules/materials, and evaluation of result confidence, thereby rapidly generating candidate schemes ready for experimental validation and connecting the prediction-to-experiment pipeline.
Capabilities on the dry side (computing, data, models, inference) are becoming increasingly accessible. However, the optimal condition windows predicted by models often exceed the physical limitations of wet-side equipment. For real laboratories to stably operate this loop, the bottleneck lies not in the first three tiers, but in the fourth tier: the experimental layer. Three core bottlenecks must be resolved simultaneously to effectively close the loop.
Bottleneck 1: Execution Bottleneck
AI systems generate abstract instructions in natural language, such as "continuously evaluate the catalyst under conditions of 80°C and 10 MPa." Laboratory instruments cannot directly interpret natural language; therefore, human intervention is required to translate these instructions into equipment actions. If manual involvement is necessary at this stage, the automated closed loop is interrupted.
Bottleneck 2: Protocol Bottleneck
A complete experiment typically involves multiple stages, including reagent preparation, reaction, post-processing, and detection, which require the coordination of several instruments. Instruments from different vendors often use incompatible communication protocols and data formats. Without device interconnection, experimental workflows cannot proceed automatically, and raw measurement data is lost at the interfaces.
Bottleneck 3: Data Bottleneck
The measured data generated during experiments often has gaps and cannot be fed back smoothly, creating a significant bottleneck. This deficiency arises from three key issues:
Lack of clean data: There is an insufficient supply of high-quality, fully contextualized, and reproducible measured data. This makes it difficult for models to extract reliable physicochemical rules.
Lack of negative samples: A skewed dataset distribution means models primarily learn successful reaction pathways and fail to predict failure scenarios. Consequently, they tend to generate candidate solutions that are not practically implementable.
Inconsistent labeling and mismatched data lineage: Labeling and parameter standards vary between batches and operators, disrupting the correspondence between experimental conditions and results. This severs the data traceability chain, creating data silos between screening and preparation systems. Without a continuous supply of high-quality measured data, AI cannot receive real-world feedback, leaving the closed loop incomplete.
At a more fundamental level, these three bottlenecks are interconnected and mutually reinforce each other. Consequently, establishing a functional dry-wet closed loop is not a matter of optimizing a single technical point, but rather a systems engineering challenge that necessitates the collaborative redesign of both hardware and software.
"Chemical Building Blocks": A Scene-Based Platform from Standardized Functional Units
As demonstrated, the experimental layer, within the four-tier architecture, presents the most significant implementation challenges. Many core components in this layer demand custom development, requiring substantial capital investment, long research and development cycles, and complex standardization. Even with fully developed core hardware components, this merely forms the foundation. Continuous reaction systems are highly coupled dynamic systems, and commissioning and validating a single product line typically takes years. The process know-how and engineering experience gained through continuous iteration cannot be replicated by simply assembling automation hardware.
Ou Shisheng's approach is to construct a comprehensive "Chemical Building Blocks" system. This system provides standardized functional modules for precise reagent dosing, temperature-controlled reactions, solid-liquid separation, sample collection, and online analysis. Each "building block" adheres to unified specifications in four key areas: mechanical interfaces, electrical interfaces, communication protocols, and data formats. This standardization allows them to be assembled into complete application platforms customized for various R&D scenarios.
This system has evolved over more than a decade of continuous iteration and technological development. It progressed from individual functional modules to integrated units, and then to complete platforms, ultimately enabling fully self-driven, end-to-end laboratories. To date, nearly one hundred standardized hardware components have been developed, encompassing external energy fields such as high/low temperature, high pressure, strong corrosion, photochemistry, and electrochemistry. This allows for the rapid, standardized delivery of application platforms for diverse scenarios.In terms of reaction types, the system supports a wide range, including transition-metal-catalyzed cross-coupling reactions (Suzuki, Heck, Negishi), redox reactions, nucleophilic substitution, condensation, cycloaddition, low-temperature organometallic reactions (lithiation, Grignard, zincation), photochemical and electrochemical synthesis, as well as oligopolyamide and polyether polymerization. For energy and catalysis applications, supported reactions include hydrogenation, selective oxidation, hydrocarbon cracking, isomerization, reforming, Fischer-Tropsch synthesis, and CO₂ conversion. The applicable scope can be summarized as follows: excluding currently unsupported systems like high-viscosity polymer systems and fluorochemical processes involving hydrogen fluoride, the system can accommodate approximately 95% of reaction types with corresponding platforms.
Figure 2 | Ou Shisheng Chemical Building Blocks
The "Brain" of the Closed Loop: Three-Tier Software Architecture for Intelligent Invocation of Building Blocks
With the "Chemical Building Blocks" fully established, software becomes the critical hub for AI integration and intelligent experimentation, acting as the "core brain" of the entire system. It performs five key functions:
Device Control: The software centrally orchestrates all hardware building blocks connected to the system, transforming complex cross-device and cross-unit workflows into automatically executable process sequences.
Protocol Unification: It abstracts hardware differences at the communication protocol and data format levels, ensuring all building blocks communicate effectively. This is the essential technical requirement for the seamless assembly of building blocks.
AI Data Upload: Setpoints and measured experimental values are automatically correlated and archived in a structured format, ready for direct upload to the AI model without any manual transcription.
Labeling and Classification: Before being fed to the model, data retains complete process context, including raw material input, operating conditions, process records, and results. It is automatically labeled and classified, transforming it into model-ready data assets rather than isolated figures.
AI Instruction Decomposition and Execution: Abstract instructions generated by the AI, such as a set of candidate reaction conditions, are automatically broken down into executable parameter sequences for each device and distributed to the hardware. The execution results are then fed back to the AI.
Figure 3 | AI4S Platform Architecture
From an architectural standpoint, this system operates on three tiers. It interfaces upward with the model layer, which includes general large models, domain-specific models, and computational models. The research intelligence assistant, Ora, orchestrates various Agents and skills, translating algorithmic recommendations into executable experimental tasks for the hardware. It connects downward to the hardware layer, managing, controlling, and recording all stages from dosing and synthesis/preparation to online detection, post-processing, material management, analytical characterization, and material transfer. Horizontally, it supports the data layer: full-process data covering experimental planning, execution, and testing evaluation is archived under unified metadata specifications, enabling intelligent invocation of the building blocks. At this stage, the complete closed-loop circuit is established.
The Dry-Wet Closed Loop: Operational in Real Laboratories, Not Just a Blueprint
This system is more than just a theoretical concept; it is actively deployed by Ou Shisheng in universities, research institutions, and industrial facilities. Ou Shisheng's closed-loop synthesis equipment has been commercially launched across various sectors, including vital industries such as pharmaceuticals, fine chemicals, catalysts, new energy and semiconductor.
Small-Molecule Pharmaceuticals and Organic Synthesis
Milligram-Scale High-Throughput Fully Automated Synthesis Workstation
This workstation performs 24-channel parallel reactions, featuring positive-pressure membrane solid-liquid separation, Rf value determination, parallel concentration, product confirmation, and inkjet marking with barcode scanning. The entire workflow, from dosing and reaction to separation, purification, concentration, and collection, runs unattended and automatically generates structured reports.
Gram-Scale High-Throughput Synthesis Workstation
This workstation allows for linear scale-up of milligram-scale screening results without the need to redevelop process parameters. Its complete workflow includes solution preparation and reaction, extraction and filtration, drying and scraping, column chromatography, concentration, weighing, and NMR sample preparation. Each process step is flexibly adjustable, supporting non-linear workflow routing and backtracking. Users can return to intermediate stages for iterative optimization at any time, significantly reducing the cycle from synthesis to the delivery of qualified samples.
Small-Molecule Synthesis Platform
This platform separates dosing and synthesis from post-processing into two independent automated workstations. It enables high-throughput parallel synthesis under anhydrous and oxygen-free conditions, along with continuous sample post-processing. This setup supports the rapid preparation of large libraries of small-molecule candidate compounds.
Materials Science
Oligomer Synthesis Platform
This platform integrates dosing, vacuum filtration, centrifugation, concentration, and in-line detection for systems with varying viscosities. Its target applications include optoelectronic display materials, photoresists, resins, and modified polymers.
High-Throughput Photochemical Synthesis System under Inert Atmosphere
Utilizing a tandem dual-glovebox architecture, this system allows flexible switching among multiple light sources. Solution preparation, reaction, and sampling occur continuously under an inert atmosphere. It supports the synthesis of pharmaceutical intermediates, fine chemicals, medicinal scaffold molecules, and organic optoelectronic materials.
Electrochemical Synthesis Platform
Driven by electric fields, this platform achieves efficient transformations under mild conditions through electron transfer. Target applications are in fuel cells, nanomaterials, two-dimensional materials, and electrochemical water treatment.
Perovskite Synthesis Platform
This platform covers the complete workflow for perovskites, including sol-gel and co-precipitation preparation, aging, drying, calcination, and photocatalytic performance evaluation. It is specifically designed for perovskites, carbon-based composite materials, and nano-photocatalytic materials.
Energy and Catalysis
Unmanned Laboratory for Catalyst Synthesis
The complete workflow, from catalyst preparation using impregnation and co-precipitation methods to performance evaluation and online analysis, operates autonomously. Notably, its fully automated catalyst evaluation unit addresses a domestic technical gap in this field, effectively replacing high-end imported equipment.
Our leading clients span diverse industries, including top pharmaceutical companies like WuXi AppTec and Pharmaron, semiconductor display enterprises such as BOE, and global chemical giants like BASF and Evonik. We also serve energy and chemical institutions within China's petroleum and petrochemical sector. This broad adoption across industries confirms the universal value of our high-throughput synthesis solution.
Data with Provenance: The Closed Loop Accumulates Not Only Experimental Results, but Valuable Assets
As previously noted, AI requires high-quality data with clear provenance: it must be clean, include negative samples, and be continuously traceable. The data generated from the dry-wet closed-loop experiments described earlier perfectly fulfills this need for high-quality, provenanced data for AI. The entire workflow runs on a single, unified platform, eliminating systematic errors from manual operations. All data, including records from failed experiments, is automatically collected, ensuring that negative samples are no longer lost. This allows the AI model to learn not only effective reaction pathways but also the boundaries of failure. Experimental operating conditions are automatically linked to raw spectra, establishing complete data provenance and preventing the disorder that often arises from manual transcription.
More importantly, because all platforms share the same software architecture and adhere to identical data specifications, experimental data generated across different scenarios and instruments are inherently consistent. This means they have uniform fields, unified labeling of process context, and consistent data lineage. This consistency enables data aggregation, horizontal comparison, and combined modeling across platforms. Each new platform deployed and each completed experiment incrementally enriches the data assets, effectively bridging previously existing data silos. The closed loop not only supports one-off research tasks but also builds a research infrastructure that grows in value over time.
Perspective: Expanding Dry-Wet Closed Loops to Leading Research Laboratories
As AI4S has progressed to its current stage, the rapid iteration of the upper three tiers is evident. However, the experimental layer truly dictates the pace of advancement in this field. Crucial questions remain: Do instruments support programmatic control? Can process data meet standardization requirements and be fully captured? Can acquired data be effectively utilized by upstream AI models? And can experimental results provide feedback to AI models, enabling automatic decision optimization for the next iteration, ultimately transforming standalone instruments into an AI-driven scientific system?
Ou Shisheng is leading the way in this competitive sector. Moving forward, Ou Shisheng will implement its dry-wet closed-loop capabilities in the R&D frontline of more universities, research institutes, and enterprises, aiming to establish a new paradigm for chemical research and development.
From the perspective of the chemical R&D industry, AI4S is fundamentally reshaping experimental research workflows. The availability of mature, domestically produced intelligent equipment provides the physical foundation for this new R&D paradigm. An end-to-end closed-loop capability that integrates both software and hardware remains a rarity in this field—and Ou Shisheng has already made this closed loop operational.
