AI in Chemicals: Why PID Autonomous Control Trumps Molecule Discovery

Introduction: Two Divergent Paths for AI in Chemical Engineering

The chemical and new materials sector operates under the strictest safety boundaries of any process industry: continuous production, high temperature and pressure, flammable and explosive materials. A single misjudgment carries far higher costs than in other sectors. Consequently, AI adoption here follows two fundamentally different trajectories:

Path 1 – Formula R&D (Incremental Innovation): Discovering new molecules and formulations. This path resembles scientific discovery: AI mines literature and patents, generates candidate molecules, robots conduct automated experiments, and results feed back into model retraining.

Path 2 – Process Control (Stock Optimization): Optimizing existing production units. This path sits closest to control authority and demands rigorous safety guarantees. It progresses from real-time sensing to parameter prediction, process optimization, closed-loop execution, and ultimately autonomous operation.

Both paths share a common foundation: predictive capability —learning the input-output relationships of complex chemical systems before attempting optimization or control.

Core Pain Point: Not Data Scarcity, But Comparability

Contrary to common belief, chemical plants are drowning in data. Lanzhou Petrochemical's Yulin smart factory built a data lake with over 200 million records, processing nearly 30,000 data points per minute, with millisecond-level synchronization across 305 critical equipment vibration parameters. The real bottleneck is cross-unit, cross-batch comparability : the same recipe yields different results when equipment or raw material batches change. This variability stems from the coupling of six-plus variables—temperature, pressure, humidity, feedstock properties, time-series dynamics, and equipment settings.

The sector faces a dual dilemma: R&D confronts combinatorial explosion (e.g., choosing 5 metals from 90 yields >43 million combinations; at 10 samples/day, exhaustive screening would take 12,040 years), while process control wrestles with multi-variable coupling where a single error can trigger catastrophic failure. Hence, AI must first solve comparability before any prediction becomes reliable.

Two Paths, One Predictive Foundation

Path 1: Formula R&D – Incremental Innovation

Workflow: Literature/patent mining → Candidate generation → Robotic automated experimentation → Data feedback for retraining.

Wanhua Chemical Catalyst Screening: Faced with 14,000+ candidate formulations, AI narrowed to 156 promising options, then to 4, enabling precise molecular synthesis recommendations.

Tianjin University & Southwest Jiaotong University: Combined LLM + genetic algorithm + ultra-fast experimentation to identify optimal catalyst in just 4 iterations and 24 samples.

TianGong Smart Materials (HENaMat): Used Crystal Graph Convolutional Neural Networks (CGCNN) to predict band gap, voltage, and capacity of sodium-ion battery anode materials with >90% accuracy (target 95%), reducing single-material computation from thousands of hours to minutes—a ~3,000× efficiency gain.

USTC "Robot Chemist": 110 robots across 19 distributed labs performing 2,000 precise operations daily. For Mars oxygen catalyst development, it absorbed 50,000+ papers, searched 3.76 million combinations (human team would need 2,000 years), and found the optimal recipe in 6 weeks.

Path 2: Process Control – Stock Optimization

Workflow: Real-time perception → Parameter prediction → Process optimization → Closed-loop execution → Autonomous operation.

Zhongkong TPT Time-Series Foundation Model at Wanhua Chlor-Alkali (Ningbo): 650k-ton caustic soda unit: pH neutralization time cut from 5 hours to 1 hour; sodium carbonate dosing precision ±0.02 g/L; electrolyzer energy consumption reduced 5%; ion membrane life prediction accuracy 95%; annual cost savings >¥10 million.

Sinopec Atmospheric & Vacuum Distillation Model: Minute-level prediction of 16 key product quality parameters with >96% accuracy; automatic optimization of 20 critical process indicators.

Lanzhou Petrochemical Yulin (China's First Full-Chain Intelligent Ethylene Plant): Ethylene yield rose from 76.2% (2021) to 78.83%; comprehensive energy consumption dropped 13%; predictive maintenance coverage reached 80%; cumulative value creation >¥10 billion.

BASF Ludwigshafen: AI process optimization lifted ethylene production efficiency 12% and cut energy use 8%.

Control Authority: A Gradual Handover Ladder

The transition from human-operated to AI-operated is not a single leap but a staged handover, each rung requiring a clear answer to "Who bears the cost of a mistake?"

Level 1: Manual operation – entirely experience-driven.

Level 2: AI advises, human-in-the-loop confirms.

Level 3: AI closed-loop control, human supervises with fallback.

Level 4: PID autonomous decision-making – AI becomes the operator.

Real-world milestones: San Ning Chemical's sulfuric acid unit reduced manual interventions from 1,600+/day to <10/day (>99% reduction), earning a spot in Hubei's first "Unmanned Factory" batch. Wanhua's Penglai Park "Unmanned Scheduling" system dynamically orchestrates 14 production units and 31 media types, achieving self-monitoring, self-intervention, and self-optimization. The prerequisite for every step is not smarter algorithms, but quantified, enforceable safety boundaries .

Evaluating with the "AI + Industry Five Questions" Framework

The series applies a consistent lens across 20 episodes:

Data Source: R&D relies on literature, patents, experimental data (USTC spent three years curating millions of chemical records); Control relies on DCS time-series and sensor logs. Data volume is abundant; comparability is scarce.

Decision Authority: R&D – humans set direction, AI proposes candidates, robots execute. Control – authority transfers incrementally from AI advice to PID autonomy. The pace of transfer determines how far a project can go.

Accountability: Chemical safety demands extreme rigor. As control shifts toward "operator" status, liability definition becomes critical. San Ning's <10 interventions/day implies robust safety interlocks and fallback mechanisms, not merely handing switches to a model.

Metrics: R&D measures experimental sample count (43M→24), compute efficiency (~3,000×), cycle time (2,000 years→6 weeks). Control tracks manual operation frequency (1,600+→<10), yield (76.2%→78.83%), energy reduction (3–13%), annual value (tens of millions RMB).

Boundaries: Formula R&D is incremental; process control is stock optimization. Their data, organizational structures, and ROI horizons differ entirely. Lab recipe ≠ successful scale-up; PID handover requires quantifiable, defensible safety envelopes.

Case Study Scorecard

Wanhua Catalyst Screening – Formula R&D – 14,000+ → 156 → 4 candidates; precise molecular synthesis recommendation – Flow Industry Review 2025

TianGong HENaMat (Na-ion anode) – Formula R&D – Band gap/voltage/capacity prediction >90%; compute: thousands of hours → minutes (~3,000×) – Xinhua 2025-12

USTC Robot Chemist – Formula R&D – 110 robots/19 labs; 2,000 ops/day; 3.76M combos solved in 6 weeks (vs. 2,000 years human) – Xinhua 2025-11

Zhongkong TPT · Wanhua Chlor-Alkali – Process Control – pH 5h→1h; Na₂CO₃ ±0.02 g/L; electrolyzer -5% energy; membrane life prediction 95%; >¥10M/yr savings – Zhongkong Tech / Baidu Baike

Wanhua Penglai "Unmanned Scheduling" / San Ning Chemical – Process Control – 14 units + 31 media real-time scheduling; San Ning ops 1,600+/day → <10/day (-99%) – Flow Industry Review 2025

Sinopec CDU/VDU / Lanzhou Yulin – Process Control – 16 quality params minute-level >96% accuracy; 20 indicators auto-optimized; Yulin ethylene yield 76.2%→78.83%, energy -13% – Industry Review / Lanzhou Petrochemical

Industry Signal: PID Autonomy Enters Official Pilot Scenarios

At the 2026 Smart Expo, the National AI Application Pilot Base (Manufacturing·Petrochemical) released the Petrochemical General Model 1.0, General Chemical Knowledge Base, and a suite of intelligent agents including PID autonomous decision-making and production planning assistants. Jointly launched by Zhejiang MIIT, Ningbo Municipal Government, and China Petroleum and Chemical Industry Federation, this inclusion of PID autonomy in the official pilot catalog marks a watershed: the capability has moved beyond lab experiments into nationally recognized, promotable scenarios.

Three Romanticized Misconceptions

Data isn't missing; comparability is. Batch consistency depends on coupled variables (temperature, humidity, feedstock, time-series, pressure, equipment settings). Without cross-unit, cross-batch comparability, models remain trapped in "single plant, single line" silos. This explains why chemical AI projects often shine in isolation but resist replication—hence the value of public goods like pilot bases and common corpora.

Control handover must quantify the cost of error first. PID autonomous decision-making means AI shifts from "advisor" to "operator" on the most conservative, core control loop in process industry. San Ning's reduction to <10 interventions/day came from safety interlocks, anomaly auto-intervention, and fallback mechanisms—not simply handing over the switch. The closer AI gets to the actuator, the more critical liability definition becomes; if that boundary isn't calculable, don't climb the next rung.

The "Lab-to-Plant" chasm is real. The Robot Chemist's 6-week discovery among 3.76M combos and Tianjin University's 24-sample catalyst identification are laboratory-scale victories. Scaling from lab recipe to 10,000-ton reactor introduces new challenges: reactor hydrodynamics, heat transfer, impurities, batch stability. The CATL-Hongziwei joint lab draws attention not just for full automation, but for achieving inter-batch result variability <3%—the true threshold from "discovery" to "usable".

Counter-Intuitive Hook

The most critical focus for chemical AI is not "discovering new molecules" but PID autonomous decision-making —the most conservative, core control loop in all process industry. Once that loop loosens, AI transforms from "advisor" to "operator". When evaluating chemical AI, don't just count new formulations; ask whether it dares, and on what grounds, to take over that core loop.

Self-Checklist for Chemical/New Materials AI Projects

Distinguish "Incremental" vs. "Stock" first. Formula R&D (new molecules/recipes) and Process Control (optimize existing assets) are separate tracks—different data sources, org structures, payback horizons. Don't use R&D budget to buy control outcomes, nor apply control safety standards to gate R&D exploration.

Control authority transfers via ladder, not leap. Manual → AI Advice → AI Closed-Loop → PID Autonomous. Each rung must answer "Who pays for a mistake?" Safety interlocks and anomaly auto-intervention must be in place before advancing.

Scrutinize comparability and scale-up, not just lab metrics. Cross-unit, cross-batch data comparability is a prerequisite. A lab "6-week recipe" must connect to pilot and commercial scale-up (batch variability, purity consistency) to count as real deployment.

Chemical and new materials close the "Hard Industry" module: they bear the strictest safety boundaries and test the patience of "learn the rules first, then talk control." The paradox is the series' sharpest tension—the most conservative control loop is the first to loosen; how far it goes depends not on algorithmic flair, but on how clearly safety boundaries are calculated. Next, we shift from "furnaces and pipes" to "money and credit"—ending Hard Industry, opening Grand Circulation: AI + Finance.