Case study — Specialty chemicals
Improving reactor yield from 71% to 84%
A temperature–pressure interaction invisible to conventional analysis was suppressing yield by 13 percentage points. Machine learning found it in six weeks.
71% → 84%
Reactor yield
13 pp
Yield improvement
£520k
Annual revenue uplift
6 weeks
Data to recommendations
The challenge
The client produced a specialty intermediate used in agrochemical synthesis. Their target reactor yield was 88%; actual performance had plateaued at 71–73% for over two years, despite a process optimisation programme that had investigated temperature, residence time, catalyst loading, and feed purity.
The 15–17 percentage point gap between target and actual yield represented a significant commercial problem. At current production volumes, each percentage point of yield improvement was worth approximately £40,000 per year in additional product. The gap was costing the business over £600,000 annually in lost output.
The process team had run a series of designed experiments (DOE) over the preceding 18 months. The experiments had identified temperature and residence time as significant factors, and the process had been adjusted accordingly — but yield had not improved. Something else was limiting performance, and the conventional approach had not found it.
The data
The client's distributed control system (DCS) logged process parameters at 5-minute intervals. We extracted 26 months of DCS data, aggregated to batch level (mean, minimum, maximum, and standard deviation for each parameter over the batch duration), and merged with the batch yield records from the LIMS.
After aggregation, we had 412 complete batch records with 67 candidate variables derived from 14 raw process parameters. The yield outcome was continuous (percentage yield), which allowed us to use regression methods rather than classification.
Dataset
Methodology
Feature engineering
In addition to the standard summary statistics (mean, min, max, SD), we engineered interaction features between the variables identified as significant in the client's previous DOE work. This gave the model explicit representations of the interactions the process team had already hypothesised, as well as the freedom to discover others.
Gradient boosting regression
We trained a gradient boosting regressor (LightGBM) on 75% of the data. The model achieved an R² of 0.84 on the held-out test set, indicating that 84% of the variance in yield could be explained by the process variables — strong evidence that the data contained the answer.
SHAP interaction analysis
SHAP interaction values decompose the model's predictions into main effects and pairwise interactions. The analysis revealed that the interaction between reactor pressure (specifically, pressure variability during the addition phase) and reactor temperature was the dominant driver of yield variation — accounting for 31% of the total explained variance.
Partial dependence plots
We generated partial dependence plots for the temperature–pressure interaction to characterise the relationship quantitatively. The plots showed a clear optimal region: yield was maximised when temperature was held between 118°C and 124°C AND pressure variability (SD over the addition phase) was below 0.8 bar. Outside this region, yield fell sharply.
Key findings
The critical finding was that pressure variability during the addition phase — not mean pressure, which the process team had been monitoring — was the limiting factor. When pressure fluctuated by more than 0.8 bar (SD) during addition, yield fell by an average of 9.3 percentage points, regardless of temperature.
The pressure variability was caused by inconsistent feed rate control during the addition phase. The feed pump was operating at the upper end of its control range, producing pressure spikes that disrupted the reaction kinetics. This had not been identified in the previous DOE work because the experiments had controlled mean pressure, not pressure variability.
The interaction with temperature explained why the previous temperature optimisation had produced inconsistent results: at high pressure variability, temperature had little effect on yield. The temperature optimisation was real, but it was being masked by the pressure variability problem.
Pressure variability during the addition phase — not mean pressure — was suppressing yield by up to 9.3 percentage points. The previous DOE had controlled the wrong variable.
Implementation
Feed pump replacement
The existing feed pump was replaced with a variable-speed pump with tighter pressure control. The new pump maintained pressure variability below 0.4 bar (SD) during the addition phase — well within the identified threshold.
Addition phase monitoring
A real-time pressure variability alarm was added to the DCS, triggering if the 10-minute rolling SD exceeded 0.6 bar during the addition phase. This provided an early warning before the threshold that caused yield loss was reached.
Temperature profile optimisation
With pressure variability controlled, the temperature optimisation identified in the previous DOE was re-implemented. The combination of controlled pressure variability and optimised temperature profile produced the full yield improvement.
Outcomes
In the three months following pump replacement and process adjustment, reactor yield stabilised at 84% — a 13 percentage point improvement over the pre-project baseline. The target of 88% was not fully achieved; the remaining gap was attributed to feed purity variation that is the subject of an ongoing supplier qualification programme.
The annualised revenue uplift from the yield improvement was £520,000 at current production volumes. The pump replacement and DCS modification cost £38,000. The project delivered a return on investment of over 13× in the first year.
The client's process team noted that the pressure variability finding had been present in the DCS data throughout the two years of the previous optimisation programme — it had simply never been analysed. The data had contained the answer the whole time.
Yield below target?
If your process has been optimised by conventional methods and yield is still not where it should be, the answer is likely in your DCS data. We can find it.
Get in touch