IndPenSim-100 — Golden batch and multiway PCA

+------------------------------------------------------------------------------
| Program Name:         step_batch.py
| Program Version:      2026-09-10 18:41:38
| Date:                 2026-09-11
| Time:                 08:50:50
| Input 1 - Raw CSV:    indpensim_process.csv  (100 batches, 23 tags)
| Input 2 - Glossary:   indpensim_schema.csv
| Input 3 - Time stamp: aligned onto 100 points of batch progress
+------------------------------------------------------------------------------

In short. 100 batches were compared against a golden batch built from 90 of them. 10 sit more than 10 sigma away at their worst point, and 10 of the 10 batches with a recorded fault were caught.

1. The input data

Where it came from

Three steps, so any number in this report can be traced back to the file it was downloaded from. The links open the file or its folder on this machine.

What	File
1. Raw data, as downloaded	100_Batches_IndPenSim_V3.csv 2.6 GB the original, kept on the drive and never edited /Volumes/YN-4T/WIP/CTRL-Designer/B-CDL-Data-Analysis/A-Raw-Data/IndPenSim	open the folder
2. Program that converted it	make_indpensim.py 8 kB run this again to rebuild the analysis file /Users/yahyanazer/Dropbox/__C_2026_Work/CTRL-Designer/B-CDL-Data-Analysis/B-Engines	open the folder
3. Data used for this analysis	indpensim_process.csv 19 MB what the numbers in this report were computed from /Users/yahyanazer/Dropbox/__C_2026_Work/CTRL-Designer/B-CDL-Data-Analysis/A-Raw-Data/IndPenSim	open the folder
Batch list	indpensim_batches.csv 5 kB /Users/yahyanazer/Dropbox/__C_2026_Work/CTRL-Designer/B-CDL-Data-Analysis/A-Raw-Data/IndPenSim	open the folder
Glossary	indpensim_schema.csv 1 kB /Users/yahyanazer/Dropbox/__C_2026_Work/CTRL-Designer/B-CDL-Data-Analysis/A-Raw-Data/IndPenSim	open the folder
Data sheet	data-source.md 8 kB what the dataset is, its licence, and its known quirks /Users/yahyanazer/Dropbox/__C_2026_Work/CTRL-Designer/B-CDL-Data-Analysis/A-Raw-Data/IndPenSim	open the folder

What it holds

Batches	100
Tags used	23
Batch length	167 to 290 h (median 230 h)
Alignment	every batch stretched onto 100 points of 0–100% progress, because they run to different lengths
Corridor	mean ± 3 sigma of the reference batches

How the batches were run

Strategy	Batches	Mean length	Mean yield
APC with Raman	30	225 h	3,484,283 kg
faulty	10	230 h	2,600,203 kg
operator controlled	30	229 h	2,833,113 kg
recipe driven	30	229 h	2,912,750 kg

Left out of the analysis: agitator_rpm (constant in every batch), paa_conc_offline (offline, 98% blank), nh3_conc_offline (offline, 98% blank), penicillin_offline (offline, 98% blank), biomass_offline (offline, 98% blank), nh3_shots (constant in every batch), viscosity_offline (offline, 98% blank).

2. The golden batch

Built from 90 reference batches. At every one of the 100 points of progress it holds a mean and a spread for each of the 23 tags. Those 90 batches sit on average 1.5% outside their own corridor, which is the noise floor of the method — nothing can score better than that.

The corridor for every tag is drawn in section 6. Here is what the whole set looks like measured against it:

the same as a picture (for printing)

One point per batch. T² across is movement inside the pattern the good batches share; SPE up is what the model cannot explain at all. Both on log scales.

3. Every batch against the golden batch

Each batch is placed by its worst single point — how far from the golden batch it ever got, in sigma. That is the plain reading of "stays within ± so many sigma". The sustained figure beside it is the 99.9th percentile, which ignores one bad sample and says where the batch really lived.

	Band	Batches	How far	Which
1	Near the golden batch — within 2 sigma	0	none
2	Similar — 2 to 5 sigma a normal batch: it wandered, but never far	27 0 with a recorded fault	worst point 2–5 sigma sustained 3 sigma (median)	Batch 2 - recipe driven, Batch 4 - recipe driven, Batch 8 - recipe driven, Batch 10 - recipe driven, Batch 11 - recipe driven, Batch 12 - recipe driven, Batch 14 - recipe driven, Batch 16 - recipe driven, Batch 22 - recipe driven, Batch 24 - recipe driven, Batch 26 - recipe driven, Batch 36 - operator controlled, Batch 38 - operator controlled, Batch 40 - operator controlled, Batch 46 - operator controlled, Batch 52 - operator controlled, Batch 56 - operator controlled, Batch 63 - APC with Raman, Batch 64 - APC with Raman, Batch 72 - APC with Raman, Batch 74 - APC with Raman, Batch 78 - APC with Raman …
3	Outside — 5 to 10 sigma clearly different from the golden batch at some point	63 0 with a recorded fault	worst point 5–9 sigma sustained 6 sigma (median)	Batch 1 - recipe driven, Batch 3 - recipe driven, Batch 5 - recipe driven, Batch 6 - recipe driven, Batch 7 - recipe driven, Batch 9 - recipe driven, Batch 13 - recipe driven, Batch 15 - recipe driven, Batch 17 - recipe driven, Batch 18 - recipe driven, Batch 19 - recipe driven, Batch 20 - recipe driven, Batch 21 - recipe driven, Batch 23 - recipe driven, Batch 25 - recipe driven, Batch 27 - recipe driven, Batch 28 - recipe driven, Batch 29 - recipe driven, Batch 30 - recipe driven, Batch 31 - operator controlled, Batch 32 - operator controlled, Batch 33 - operator controlled …
4	Far from it — over 10 sigma a different batch altogether - something happened	10 10 with a recorded fault	worst point 14–260 sigma sustained 39 sigma (median)	Batch 91 - faulty, Batch 92 - faulty, Batch 93 - faulty, Batch 94 - faulty, Batch 95 - faulty, Batch 96 - faulty, Batch 97 - faulty, Batch 98 - faulty, Batch 99 - faulty, Batch 100 - faulty

A batch holds 23 tags x 100 points = 2,300 readings. The largest of that many draws from a normal distribution is about 3.5 sigma by chance alone, so the first band is expected to be thinly populated or empty even for perfect batches. Change the edges in the window if you want bands that split your own data more usefully.

4. What caused the deviation

The 10 batches beyond 10 sigma, worst first. "Worst tag" is the single tag furthest from the golden batch, and "at" is where in the batch that happened.

Batch	Fault recorded	Worst point	Worst tag	at	first left	time outside
91	faulty	yes	260 sigma	acid_flow_rate	52%	2%	22.6%
94	faulty	yes	238 sigma	acid_flow_rate	44%	9%	10.6%
100	faulty	yes	238 sigma	acid_flow_rate	44%	6%	5.9%
95	faulty	yes	179 sigma	acid_flow_rate	48%	10%	13.3%
99	faulty	yes	128 sigma	substrate_conc	46%	2%	15.5%
92	faulty	yes	76 sigma	ph	69%	7%	1.1%
98	faulty	yes	67 sigma	ph	69%	13%	0.7%
93	faulty	yes	40 sigma	temperature	42%	1%	2.0%
96	faulty	yes	26 sigma	temperature	36%	33%	1.3%
97	faulty	yes	14 sigma	sugar_feed_rate	40%	9%	2.6%

The milder bands

Similar — 2 to 5 sigma — 27 batches, 0 with a recorded fault. Furthest tag most often: temperature (8), heat_water_flow (4), paa_flow (4). The worst point typically falls at 39% of the batch.
Outside — 5 to 10 sigma — 63 batches, 0 with a recorded fault. Furthest tag most often: acid_flow_rate (29), heat_water_flow (13), dumped_broth_flow (8). The worst point typically falls at 43% of the batch. acid_flow_rate is the furthest tag in 29 of these 63 batches. When one tag leads across a whole band of otherwise normal batches, the corridor is too tight for that tag rather than the batches being wrong — it swings more than the 3 sigma corridor allows even when nothing is amiss.

Batch by batch

Each chart holds every tag of that batch on one axis, measured in sigma from the golden batch. The green line at zero is the golden trajectory, the dashed red lines are the ±3 sigma corridor. The tags driving the batch are bold, the rest are faint behind them so you can still see whether the whole batch moved or only a few tags.

1. Batch 91 - faulty does not look like a golden batchseverity 10.0
faulty · a fault IS recorded for this batch · first left the corridor at 2% of the way through

the same as a picture (for printing)

What the chart shows. substrate_conc carries most of it, running 249 sigma above the golden batch at 45% of the way through; 20 of 23 tags leave the corridor at some point; the first departure is at 2%, almost from the start; a large share of the batch sits outside, so this batch did not recover.

The numbers. SPE 311632 against a limit of 337 (925.4x); T² 435 against a limit of 87 (5.0x); 22.6% of the batch sits outside the golden corridor.

Driven by: substrate_conc (66%), acid_flow_rate (30%), dumped_broth_flow (1%), sugar_feed_rate (1%), temperature (0%)

2. Batch 94 - faulty does not look like a golden batchseverity 10.0
faulty · a fault IS recorded for this batch · first left the corridor at 9% of the way through

the same as a picture (for printing)

What the chart shows. acid_flow_rate carries most of it, running 238 sigma above the golden batch at 44% of the way through; 13 of 23 tags leave the corridor at some point; the first departure is at 9%, almost from the start.

The numbers. SPE 77354 against a limit of 337 (229.7x); T² 237 against a limit of 87 (2.7x); 10.6% of the batch sits outside the golden corridor.

Driven by: acid_flow_rate (74%), temperature (7%), co2_offgas (6%), cool_water_flow (4%), substrate_conc (3%)

3. Batch 100 - faulty does not look like a golden batchseverity 10.0
faulty · a fault IS recorded for this batch · first left the corridor at 6% of the way through

the same as a picture (for printing)

What the chart shows. acid_flow_rate carries most of it, running 238 sigma above the golden batch at 44% of the way through; 15 of 23 tags leave the corridor at some point; the first departure is at 6%, almost from the start.

The numbers. SPE 75142 against a limit of 337 (223.1x); T² 93 against a limit of 87 (1.1x); 5.9% of the batch sits outside the golden corridor.

Driven by: acid_flow_rate (76%), co2_offgas (9%), substrate_conc (7%), o2_offgas (3%), aeration_rate (1%)

4. Batch 95 - faulty does not look like a golden batchseverity 10.0
faulty · a fault IS recorded for this batch · first left the corridor at 10% of the way through

the same as a picture (for printing)

What the chart shows. acid_flow_rate carries most of it, running 179 sigma above the golden batch at 48% of the way through; 16 of 23 tags leave the corridor at some point; the first departure is at 10%, early in the batch.

The numbers. SPE 60438 against a limit of 337 (179.5x); T² 190 against a limit of 87 (2.2x); 13.3% of the batch sits outside the golden corridor.

Driven by: acid_flow_rate (92%), sugar_feed_rate (3%), ph (1%), temperature (1%), cool_water_flow (1%)

5. Batch 99 - faulty does not look like a golden batchseverity 10.0
faulty · a fault IS recorded for this batch · first left the corridor at 2% of the way through

the same as a picture (for printing)

What the chart shows. substrate_conc carries most of it, running 128 sigma above the golden batch at 45% of the way through; 21 of 23 tags leave the corridor at some point; the first departure is at 2%, almost from the start; a large share of the batch sits outside, so this batch did not recover.

The numbers. SPE 59809 against a limit of 337 (177.6x); T² 260 against a limit of 87 (3.0x); 15.5% of the batch sits outside the golden corridor.

Driven by: substrate_conc (72%), co2_offgas (7%), cool_water_flow (4%), heat_water_flow (4%), temperature (3%)

6. Batch 92 - faulty does not look like a golden batchseverity 10.0
faulty · a fault IS recorded for this batch · first left the corridor at 7% of the way through

the same as a picture (for printing)

What the chart shows. ph carries most of it, running 76 sigma below the golden batch at 69% of the way through; 6 of 23 tags leave the corridor at some point; the first departure is at 7%, almost from the start; only a small part of the batch is outside, so this is a short, sharp deviation rather than a batch that ran wrong throughout.

The numbers. SPE 12162 against a limit of 337 (36.1x); 1.1% of the batch sits outside the golden corridor.

Driven by: ph (91%), base_flow_rate (2%), temperature (1%), heat_water_flow (1%), acid_flow_rate (1%)

7. Batch 98 - faulty does not look like a golden batchseverity 10.0
faulty · a fault IS recorded for this batch · first left the corridor at 13% of the way through

the same as a picture (for printing)

What the chart shows. ph carries most of it, running 67 sigma below the golden batch at 69% of the way through; 4 of 23 tags leave the corridor at some point; the first departure is at 13%, early in the batch; only a small part of the batch is outside, so this is a short, sharp deviation rather than a batch that ran wrong throughout.

The numbers. SPE 8936 against a limit of 337 (26.5x); 0.7% of the batch sits outside the golden corridor.

Driven by: ph (91%), heat_water_flow (2%), base_flow_rate (1%), temperature (1%), cool_water_flow (1%)

8. Batch 93 - faulty does not look like a golden batchseverity 10.0
faulty · a fault IS recorded for this batch · first left the corridor at 1% of the way through

the same as a picture (for printing)

What the chart shows. temperature carries most of it, running 40 sigma above the golden batch at 42% of the way through; 10 of 23 tags leave the corridor at some point; the first departure is at 1%, almost from the start; only a small part of the batch is outside, so this is a short, sharp deviation rather than a batch that ran wrong throughout.

The numbers. SPE 6152 against a limit of 337 (18.3x); 2.0% of the batch sits outside the golden corridor.

Driven by: temperature (74%), generated_heat (6%), cool_water_flow (5%), dumped_broth_flow (2%), base_flow_rate (2%)

9. Batch 96 - faulty does not look like a golden batchseverity 10.0
faulty · a fault IS recorded for this batch · first left the corridor at 33% of the way through

the same as a picture (for printing)

What the chart shows. temperature carries most of it, running 26 sigma above the golden batch at 36% of the way through; 4 of 23 tags leave the corridor at some point; the first departure is at 33%, around the middle; only a small part of the batch is outside, so this is a short, sharp deviation rather than a batch that ran wrong throughout.

The numbers. SPE 2986 against a limit of 337 (8.9x); 1.3% of the batch sits outside the golden corridor.

Driven by: temperature (65%), cool_water_flow (13%), generated_heat (8%), heat_water_flow (4%), ph (1%)

10. Batch 97 - faulty does not look like a golden batchseverity 9.8
faulty · a fault IS recorded for this batch · first left the corridor at 9% of the way through

the same as a picture (for printing)

What the chart shows. sugar_feed_rate carries most of it, running 14 sigma below the golden batch at 40% of the way through; 12 of 23 tags leave the corridor at some point; the first departure is at 9%, almost from the start; only a small part of the batch is outside, so this is a short, sharp deviation rather than a batch that ran wrong throughout.

The numbers. SPE 2467 against a limit of 337 (7.3x); 2.6% of the batch sits outside the golden corridor.

Driven by: sugar_feed_rate (40%), cool_water_flow (11%), base_flow_rate (8%), temperature (7%), acid_flow_rate (6%)

5. How each way of running the plant behaves

Each curve is one batch: how far it sits from the golden batch at that point of its progress, taken across all 23 tags at once. The bold line is the middle batch of the group. All four charts share the same scale, so they can be read against each other.

Way of running it	Batches	Typical distance from golden	Worst batch	Mean yield	Beyond 10 sigma
recipe driven	30	0.71 sigma	9 sigma	2,912,750 kg	0 of 30
operator controlled	30	0.74 sigma	9 sigma	2,833,113 kg	0 of 30
APC with Raman	30	0.63 sigma	9 sigma	3,484,283 kg	0 of 30
faulty	10	1.12 sigma	260 sigma	2,600,203 kg	10 of 10

The four side by side

Only the middle batch of each group, so the four are directly comparable.

the same as a picture (for printing)

recipe driven — 30 batches

the same as a picture (for printing)

operator controlled — 30 batches

the same as a picture (for printing)

APC with Raman — 30 batches

the same as a picture (for printing)

faulty — 10 batches

the same as a picture (for printing)

6. Every batch, ranked

label	batch	strategy	band	max_z	sustained_z	worst_tag	worst_at_pct	outside_pct	first_outside_pct	T2	SPE	yield_total_kg
Batch 91 - faulty	91	faulty	far from it (over 10 sigma)	260.2	214.1	acid_flow_rate	51.5	22.57	2	435.4	3.116e+05	2.597e+06
Batch 99 - faulty	99	faulty	far from it (over 10 sigma)	128.3	84.2	substrate_conc	45.5	15.52	2	259.7	5.981e+04	8.908e+05
Batch 45 - operator controlled	45	operator controlled	outside (5-10 sigma)	9.4	9.4	sugar_feed_rate	48.5	13.52	2	86.52	7.6	2.828e+06
Batch 95 - faulty	95	faulty	far from it (over 10 sigma)	179.3	40.5	acid_flow_rate	48.5	13.35	10.1	190.4	6.044e+04	1.954e+06
Batch 67 - APC with Raman	67	APC with Raman	outside (5-10 sigma)	9.4	9.4	head_pressure	7.1	13.17	2	87.64	2.188	3.413e+06
Batch 94 - faulty	94	faulty	far from it (over 10 sigma)	237.8	44.6	acid_flow_rate	44.4	10.57	9.1	237.1	7.735e+04	3.584e+06
Batch 29 - recipe driven	29	recipe driven	outside (5-10 sigma)	9.4	9.4	aeration_rate	14.1	9.13	1	86.66	8.515	1.842e+06
Batch 3 - recipe driven	3	recipe driven	outside (5-10 sigma)	9.4	9.2	acid_flow_rate	9.1	8.87	1	86.8	5.971	2.675e+06
Batch 9 - recipe driven	9	recipe driven	outside (5-10 sigma)	9.4	9.4	acid_flow_rate	47.5	7.43	2	80.8	45.04	4.083e+06
Batch 73 - APC with Raman	73	APC with Raman	outside (5-10 sigma)	9.4	8.9	acid_flow_rate	34.3	6.61	5.1	78.66	55.14	3.549e+06
Batch 100 - faulty	100	faulty	far from it (over 10 sigma)	237.8	42.5	acid_flow_rate	44.4	5.87	6.1	93.1	7.514e+04	1.555e+06
Batch 15 - recipe driven	15	recipe driven	outside (5-10 sigma)	9.4	7.3	substrate_conc	8.1	4.3	2	53.92	237.9	2.853e+06
Batch 5 - recipe driven	5	recipe driven	outside (5-10 sigma)	9.4	6.6	sugar_feed_rate	44.4	3.57	2	48.82	275.8	3.563e+06
Batch 75 - APC with Raman	75	APC with Raman	outside (5-10 sigma)	9.4	9.1	acid_flow_rate	8.1	3.39	4	82.58	30.17	3.1e+06
Batch 44 - operator controlled	44	operator controlled	outside (5-10 sigma)	7.9	5.5	heat_water_flow	39.4	3.35	5.1	47.68	205.5	1.474e+06
Batch 61 - APC with Raman	61	APC with Raman	outside (5-10 sigma)	9	7.1	ph	34.3	3.13	4	56.8	184.5	3.063e+06
Batch 35 - operator controlled	35	operator controlled	outside (5-10 sigma)	9.4	8.2	dumped_broth_flow	51.5	2.91	1	79.92	48.45	3.538e+06
Batch 17 - recipe driven	17	recipe driven	outside (5-10 sigma)	9.4	6.8	acid_flow_rate	53.5	2.91	2	81.51	33.23	3.852e+06
Batch 25 - recipe driven	25	recipe driven	outside (5-10 sigma)	7.9	6.2	acid_flow_rate	97	2.74	2	48.54	215.9	3.717e+06
Batch 97 - faulty	97	faulty	far from it (over 10 sigma)	14.4	11.6	sugar_feed_rate	40.4	2.61	9.1	34.29	2467	3.894e+06
Batch 33 - operator controlled	33	operator controlled	outside (5-10 sigma)	9.4	8.6	acid_flow_rate	81.8	2.48	2	74.36	81.72	3.294e+06
Batch 23 - recipe driven	23	recipe driven	outside (5-10 sigma)	7.4	5.3	acid_flow_rate	49.5	2.35	2	60.57	167.1	2.855e+06
Batch 7 - recipe driven	7	recipe driven	outside (5-10 sigma)	9.3	9.2	ph	77.8	2.13	43.4	83.26	31.42	3.455e+06
Batch 93 - faulty	93	faulty	far from it (over 10 sigma)	39.9	32	temperature	42.4	2.04	1	61.7	6152	4.448e+06
Batch 39 - operator controlled	39	operator controlled	outside (5-10 sigma)	9.4	9.4	acid_flow_rate	99	2	18.2	83.1	26.98	2.682e+06
Batch 32 - operator controlled	32	operator controlled	outside (5-10 sigma)	7.6	6.1	temperature	37.4	1.96	20.2	41.17	228	4.196e+06
Batch 53 - operator controlled	53	operator controlled	outside (5-10 sigma)	5.7	5.4	heat_water_flow	13.1	1.61	1	43.01	246.6	2.858e+06
Batch 31 - operator controlled	31	operator controlled	outside (5-10 sigma)	7.5	6.4	heat_water_flow	33.3	1.52	2	67.55	139.8	1.89e+06
Batch 69 - APC with Raman	69	APC with Raman	outside (5-10 sigma)	5	4.8	dumped_broth_flow	58.6	1.48	35.4	57.29	144	3.751e+06
Batch 30 - recipe driven	30	recipe driven	outside (5-10 sigma)	9.3	4.9	acid_flow_rate	57.6	1.39	55.6	37.82	298.9	3.539e+06
Batch 96 - faulty	96	faulty	far from it (over 10 sigma)	26.4	19.1	temperature	36.4	1.3	33.3	39.52	2986	3.595e+06
Batch 59 - operator controlled	59	operator controlled	outside (5-10 sigma)	9.3	6.6	acid_flow_rate	84.8	1.3	7.1	52.07	202.6	2.91e+06
Batch 27 - recipe driven	27	recipe driven	outside (5-10 sigma)	9.4	6.6	substrate_conc	42.4	1.22	23.2	63.25	142.4	2.176e+06
Batch 41 - operator controlled	41	operator controlled	outside (5-10 sigma)	9.4	8.5	acid_flow_rate	94.9	1.09	10.1	82.72	27.41	2.922e+06
Batch 92 - faulty	92	faulty	far from it (over 10 sigma)	76.2	37.7	ph	68.7	1.09	7.1	57.81	1.216e+04	1.753e+06
Batch 71 - APC with Raman	71	APC with Raman	outside (5-10 sigma)	9.4	6.3	acid_flow_rate	79.8	1.04	6.1	59.89	167.1	2.831e+06
Batch 83 - APC with Raman	83	APC with Raman	outside (5-10 sigma)	9.4	6	acid_flow_rate	62.6	1	10.1	48.17	222.3	3.837e+06
Batch 37 - operator controlled	37	operator controlled	outside (5-10 sigma)	9.4	7	acid_flow_rate	71.7	1	12.1	51.15	209.5	2.632e+06
Batch 21 - recipe driven	21	recipe driven	outside (5-10 sigma)	6.3	5.3	heat_water_flow	18.2	0.96	5.1	52.74	195.6	2.568e+06
Batch 20 - recipe driven	20	recipe driven	outside (5-10 sigma)	8.1	6.2	heat_water_flow	46.5	0.96	36.4	48.35	197.7	2.234e+06

first 40 of 100 — the rest are in batch_scores.csv

7. The golden batch, tag by tag

The green band is where a good batch is expected to be at each point of its progress. The worst batches are drawn over it in red, and one reference batch in green for comparison. Drag to zoom, double-click to reset, click a name in the legend to hide it.

aeration_rate

the same as a picture (for printing)

sugar_feed_rate

the same as a picture (for printing)

acid_flow_rate

the same as a picture (for printing)

base_flow_rate

the same as a picture (for printing)

cool_water_flow

the same as a picture (for printing)

heat_water_flow

the same as a picture (for printing)

water_injection

the same as a picture (for printing)

head_pressure

the same as a picture (for printing)

dumped_broth_flow

the same as a picture (for printing)

substrate_conc

the same as a picture (for printing)

dissolved_o2

the same as a picture (for printing)

penicillin_conc

the same as a picture (for printing)

vessel_volume

the same as a picture (for printing)

vessel_weight

the same as a picture (for printing)

ph

the same as a picture (for printing)

temperature

the same as a picture (for printing)

generated_heat

the same as a picture (for printing)

co2_offgas

the same as a picture (for printing)

paa_flow

the same as a picture (for printing)

oil_flow

the same as a picture (for printing)

oxygen_uptake

the same as a picture (for printing)

o2_offgas

the same as a picture (for printing)

carbon_evolution

the same as a picture (for printing)

8. What this cannot tell you

Batches are aligned by percent complete. If a batch is slow because a phase ran long, alignment by phase or by an indicator variable would place it better.
The corridor is only as good as the reference batches. If a bad batch is in the reference set, its behaviour becomes normal.
The bands say how far a batch sat from the golden one. They do not say the deviation mattered: a batch inside the corridor throughout can still be poor if it is late, or if the yield is low for reasons no tag records.
"Worst tag" is where the deviation is largest, which is not always where it started. Use the per-batch charts above to see the order things moved in.

9. Comparing the ways of running the plant

APC with Raman produces the most and is also the most repeatable.
Across all 3 strategies the yields differ: one-way ANOVA F=9.91, p=1.3e-04; Kruskal-Wallis p=5.1e-04 without assuming normality. The spreads differ too: Levene p=5.6e-05.
The 10 faulty batches are excluded from every comparison below. They appear only in the last table, as the thing each strategy's golden batch is asked to detect.

9.1 What each way of running it produced

Way of running it	Batches	Mean yield (kg)	Spread (CV)	Worst batch	Best batch	Mean hours	kg per hour
recipe driven	30	2,912,750	26.4%	1,842,400	4,083,100	228.8	12,917
operator controlled	30	2,833,113	24.0%	1,474,100	4,196,000	229.2	12,431
APC with Raman	30	3,484,283	8.8%	2,830,800	4,044,300	224.7	15,587

CV is the batch-to-batch spread as a percentage of the mean. In a regulated plant it often matters more than the mean, because it is what lets you promise a delivery date and pass a validation.

9.2 Is the difference real, or is it noise?

Comparison	Difference		Welch p	Mann-Whitney p	Cohen d	Significant
recipe driven vs operator controlled	+79,637 kg	+2.8%	6.7e-01	8.9e-01	+0.11	no
recipe driven vs APC with Raman	-571,533 kg	-16.4%	5.3e-04	1.1e-02	-0.98	yes
operator controlled vs APC with Raman	-651,170 kg	-18.7%	2.3e-05	4.4e-05	-1.23	yes

Two tests because one assumes normality and the other does not; they should agree. Cohen d is the size of the gap in standard deviations: 0.2 is small, 0.5 medium, 0.8 large. A difference can be statistically significant and still too small to act on, so read the size as well as the p-value.

9.3 The golden batch of each way of running it

One real batch per strategy, chosen on two criteria at once: high yield and close to that group's own mean trajectory. The highest-yielding batch is often a lucky outlier that ran unlike anything else, and the most typical batch is often mediocre. These are both.

Way of running it	Golden batch	Yield	Yield rank	Closeness rank
recipe driven	Batch 26 - recipe driven	3,886,800 kg	4 of 30	2 of 30
operator controlled	Batch 60 - operator controlled	3,802,300 kg	3 of 30	5 of 30
APC with Raman	Batch 79 - APC with Raman	3,825,600 kg	5 of 30	2 of 30

9.4 Does each way of running it actually look different?

A golden batch is built from one strategy, holding back some of its own batches, and every group is then scored against it. The figure is the median batch's typical distance, in sigma. The shaded cell on the diagonal is that strategy's own held-back batches — the fair self-comparison, and it should be the lowest number in its row.

Golden built from	recipe driven	operator controlled	APC with Raman	faulty
recipe driven	0.77	1.07	1.08	1.11
operator controlled	0.82	0.74	0.95	1.20
APC with Raman	1.05	1.35	0.66	12.51

Golden built from	Its own held-back batches	The other strategies	How much further	Fault detection power
recipe driven	0.77	1.07	1.40x	1.4x
operator controlled	0.74	0.89	1.20x	1.6x
APC with Raman	0.66	1.20	1.82x	19.0x

"How much further" says how distinctive that way of running the plant is. "Fault detection power" is how far the faulty batches sit compared with its own good ones — a tighter normal is a more sensitive detector, which is a reason to run a strategy quite separate from what it yields.

9.5 What this comparison cannot settle

Do not rank strategies by distance from the main golden batch in section 2. That golden is built from all the good batches, so it is a blend of every strategy, and each is being measured against an average that contains itself. Section 9.4 avoids that by giving each strategy its own golden.
These are the batches that were run, not a designed experiment. If the strategies were used in different periods, on different feedstock or by different crews, that difference is inside the numbers and cannot be separated out.
Yield is the only outcome recorded here. Cost, energy, operator effort and the price of the instrumentation are not, and a strategy that yields more can still lose on all four.
Sample size is 30 to 30 batches per strategy. Enough to see a large difference, not enough to see a small one.

Folder: /Volumes/YN-4T/WIP/CTRL-Designer/B-CDL-Data-Analysis/E-Batch/IndPenSim-100