Syngenta: Steam consumption optimization

Objectives

• Develop a reliable predictive model for steam consumption based on historical process data at the chemical plant site in Kaisten, Switzerland.
• Identify operational levers influencing steam efficiency.
• Optimize distillation column feed strategies to minimize steam usage while meeting production targets.
• Provide actionable recommendations and tools to plant operators.

Approach

The project was executed in two main steps, combining data-driven modeling with prescriptive optimization.

Step 1 — Steam Prediction Model

‍We first developed a reliable simulation of column steam consumption. After exploratory analysis and unsuccessful attempts at purely physics-based simulation using PID-controller logic, we adopted an AI-based time-series forecasting model to predict short-term steam consumption from historical process signals.

Transformer-based models were trained for each distillation column, using time series of historical data (feed, water, reflux ratio, and cross-column interactions capturing heat recovery effects). The model achieved strong predictive performance on unseen data, with R² ≈ 0.95 for either column (Figure 1). To enable fast optimization, distilled neural network surrogate models were later trained to approximate the transformers with significantly lower computational cost while preserving high predictive performance.

‍

Step 2 — Optimization Model

‍Using the steam prediction model as a digital surrogate of the plant’ distillation process, we formulated an optimization problem to find feed strategies that minimize total weekly steam consumption while satisfying:
• Production targets (e.g., weekly end product demand)
• Operational constraints (feed bounds, tanks capacity limits, operators preference)

Two complementary optimization methods were evaluated:
• Linear Programming (LP) — exact optimization after linearization of the steam model.
• Genetic Algorithms (GA) — flexible statistical optimization without linearity assumptions.

The LP approach provided theoretically optimal solutions under defined constraints (Figure 2), while GA offered robust alternatives for highly non-linear scenarios.

Key Findings

• Heat recovery interactions between distillation columns significantly impact efficiency.
• Operating both columns together is generally more energy efficient.
• The AI steam model can reliably evaluate candidate operating strategies.
• Optimization reveals consistent over-steaming in historical operations, indicating real savings potential.

Impact

On representative test weeks, the optimized strategies achieved meaningful steam reductions while meeting production requirements. For example, during the example week in Figure 2 (March 3rd–10th, 2025), the optimized plan reduced steam usage from 312 tons to 268 tons,  a saving of 44 tons of steam for that week.

Across broader historical simulations, constrained optimization scenarios indicated substantial cumulative savings potential without requiring hardware changes. The work demonstrates that data-driven operational optimization can deliver measurable energy and cost benefits in continuous chemical processes.

Beyond immediate savings, the delivered steam model provides Syngenta with a reusable digital capability to:
• Evaluate future operating strategies offline
• Support operator decision-making
• Enable future closed-loop or advisory optimization systems.

Future Opportunities

• Extend modeling to product quality prediction.
• Incorporate forecasts of stochastic inputs (tank inflow, disturbances).
• Deploy decision-support tools for operators.
• Explore real-time optimization integration.

The project establishes a strong foundation for AI-driven process optimization at scale.