Sequencing semiconductor work orders with Random Forest

My master's research began with a problem I saw every day in a semiconductor plant: work orders were assigned to machines without matching each machine's capacity, causing bottlenecks, long waits, and some machines idle while others overflowed. The research question was simple: if we could predict in advance which step each work order would reach and how long it would take, how much better could we sequence the work?

This article walks through the whole approach in three phases, build, analyse, and deploy for real use, as in the drawing above.

Phase 1, build: from raw data to a usable dataset

The data came from two years of the production system's PostgreSQL database: 17,953,082 rows at the start, 4,052,934 after filtering, split into 2,837,053 training rows and 1,215,881 test rows.

Step	What was done	Why
Filter	Remove incomplete or abnormal rows	Wrong data teaches the model wrong things
Encode	Label Encoding from text to numbers	Models accept numbers only
Scale	Bring each column into the same range	Large-valued columns should not dominate
Split	70% training, 30% test	Test on data the model has never seen

All of this runs as an automated 15-step pipeline, from loading settings and querying data through cleaning, parallel training and evaluation to saving and deploying the model, so it can be re-run whenever new data arrives.

Figure: the 15-step pipeline for building and deploying the model. Source: the author (2025).

Phase 2, analyse: a fair comparison of three models

I chose three models suited to tabular data and measured them with the same metrics. Accuracy alone is not enough, because production data is imbalanced: some steps are common and some rare, so macro-averaged F1 tells the truth better.

Model	Accuracy	Precision	Recall	F1
Random Forest	99.97%	98.99%	98.81%	98.90%
Hist Gradient Boosting	96.65%	78.51%	76.69%	77.46%
Logistic Regression	79.62%	83.23%	23.70%	24.19%

Figure: metrics and run time compared. Source: the author (2025).

How to read it. Logistic Regression looks fine at 79.62% accuracy, but its F1 is only 24.19%: it is right mostly on the common classes. Random Forest scores high on both accuracy and F1, with a cross-validated mean of 99.96% and a standard deviation of just 0.0000346, so results are stable across every split. The trade-off is a longer training time than Hist Gradient Boosting, which is acceptable because training runs outside working hours.

Why Random Forest. Production data is non-linear and imbalanced. Hundreds of decision trees voting together handle that well, tolerate outliers, and can explain which features matter, which is essential when talking with engineers on the floor.

Phase 3, deploy: put it where people work

The best model is saved to a file and exposed through a Scheduler API running on Linux and Docker. The planning system sends work-order data and gets predictions back, and an Android app keeps working without signal.

The sequencing screen shows waiting orders ranked by the model next to the machines currently running. The planner still makes the final call, and whenever the real result differs from the prediction, that data goes back into the next training round.

The approach I want to share

1. Start from a measurable problem on the floor, not from an algorithm you want to try.
2. Use the data the factory already has. Two years of production history is worth more than it looks.
3. Measure fully, especially with imbalanced data, where accuracy alone misleads.
4. Make it re-runnable and automatic, so it can be retrained when the line changes.
5. Deliver it inside the systems people already use, through an API, with people making the final decision.

A caution: accuracy this high reflects a fairly stable process. When products change or machines are added, watch the accuracy and retrain regularly. Company details are confidential, so the screen in this article is a concept drawing.

References

Boonnuk, P., & Romsaiyud, W. (2027). Predictive work order management for semiconductor manufacturing processes using the random forest algorithm [in Thai]. KMUTNB Journal of Information Technology, 23(1). (Accepted for publication)