ALUMINA PROCESS ANALYTICS

Investigating abnormal
alumina decomposition

Process monitoring and factor diagnosis
with XGBoost, SHAP and directional rules

Bayer process precipitationTechnical project presentation
Al2O3
01
PROCESS BACKGROUND

The Bayer process

Bauxite is refined into alumina through four main stages.

01
Digestion		02
Clarification		03
Precipitation		04
Calcination
Hot caustic soda dissolves aluminium-bearing minerals.Undissolved impurities are separated from the liquor.Cooling and Al(OH)3 seeds promote hydrate crystal growth.Heat removes bound water to produce alumina.
Aluminate liquorClarified liquorAl(OH)3 hydrateAl2O3
2 Al(OH)3 →heat Al2O3 + 3 H2O

Project focus: abnormal performance in the precipitation section.

Sources: Australian Aluminium Council; The Aluminum Association. Chemistry: International Aluminium Institute, p. 12.
02
PROBLEM DEFINITION

Project objective

Identify candidate process drivers of abnormal decomposition efficiency and prioritize engineering investigation.

Process metric

η = 1 − αinαout
α = 1.645 × NKAO

Computed from measured inlet and outlet values.
η = 0.4666 means 46.66% efficiency.

Diagnostic challenges

  • Several process inputs change together.
  • Outlet measurements arrive after the input conditions.
  • Some solids and oxalate measurements are sparse.
The project metric is decomposition efficiency. It does not measure production throughput per hour.
03
DATASET

Dataset overview

1,139 samples37 columnsOct 2, 2025–Oct 19, 2026
Column groupWhat the columns describe
Time and work shiftTimestamp, work shift and the matched input-to-outlet interval
Tank temperaturesSeven sensors at tanks 1, 2, 6, 8, 12, 13 and 14
Inlet chemistryAO (dissolved alumina), NT (total alkali), NK (caustic alkali), SiO2 (silica), αK and inlet solids
Seed and tank solidsSeed solids and 14 individual tank columns measure solid content
OxalateOxalate concentration in the inlet liquor
Outlet chemistryMatched outlet NT, NK, AO and αK after precipitation
EfficiencyDecomposition efficiency calculated from inlet and outlet NK/AO

Measurement availability

FieldAvailableMissing
Oxalate181958
Any tank solids371768
Inlet SiO21,12910
Inlet solids1,12712
Seed solids1,1372

Counts use all 1,139 samples. “Any tank solids” requires at least one of the 14 tank measurements.

αK is the caustic-to-alumina ratio, calculated as 1.645 × NK/AO. Recorded matching interval: 46h; the outlet fields retain their original 48h labels.
04
DIAGNOSTIC TARGET

Why focus the diagnostic model on outlet AO?

Relationship to efficiency

η = 1 − NKinAOin × AOoutNKout

With inlet chemistry and outlet NK fixed, higher outlet AO means lower efficiency.

AO provides a directly measured outcome for the diagnostic model. The full efficiency calculation still uses all four chemistry values.

Formula illustration
Outlet AO 82.8 to 83.8
Efficiency 46.66% to 46.02%
Other chemistry values held fixed.
Measured outlet AO versus decomposition efficiency across 911 training observations; Pearson correlation minus 0.893
AO is the selected diagnostic branch, not a claim that other chemistry is irrelevant. The observed association partly reflects the efficiency formula and does not establish causal sufficiency.
05
SYSTEM DESIGN

Separate monitoring from explanation

ModuleMethodOutput
Abnormality monitoringMeasured outlet AO
Four SPC reference rules		Normal / warning / abnormal
Alert events and trigger reasons
Factor explanationProcess features and measured AO class
XGBoost + SHAP + directional rules		Ranked supporting factors
Values, reference ranges and weights

Independent maintenance

Alert limits or trigger rules can change without automatically retraining the explanation model.

Aligned records

Measurement identity and time alignment connect the monitoring event to the relevant process inputs.

The architecture describes two separately implemented modules. A unified production notification service is outside the demonstrated scope.
06
MONITORING

Four rules for abnormal outlet AO

RuleTriggerLevel
R1 · Large deviationAO outside μ ± 3sAbnormal
R2 · Sustained deviationAt least 2 of the latest 3 points beyond 2s on the same sideWarning
R3 · Rising trend6 consecutive points strictly increasingAbnormal
R4 · Adjacent jump|AOt − AOt−1| > 3.267 × baseline mean moving rangeWarning

μ and s are the mean and sample standard deviation of a fixed historical reference period.

When several rules trigger, use the highest alert level and retain all trigger reasons.

R2–R4 require consecutive valid 8-hour measurements. Equal boundaries do not trigger. Missing values and sampling gaps interrupt continuous patterns.
07
DATA AND FEATURES

Twelve engineered process factors

GroupModel factorsConstruction
Work scheduleWork shiftThree 8-hour work shifts: morning, afternoon and night
Tank temperaturesFirst / middle / last tank temperaturesAvailable-value means of tanks 1–2 / 6 & 8 / 12–14
Inlet liquorNC, AO, NK, SiO2, solidsNC = NT − NK; NK remains a separate input
Solids and impuritySeed solids, mean slurry solids, oxalateSlurry solids average available tanks 1–14

Current-row features only. Training references stay fixed during inference.

Temperature and tank-solids averages use available measurements. NC requires both NT and NK. Missing inputs are not replaced by zero; outlet measurements and efficiency are excluded from inputs.
08
INPUT ANALYSIS

Correlation among the process factors

Spearman correlation heatmap of eleven numerical process factors, with coefficients from minus one to one
911 training records. Pairwise available Spearman correlations; valid pairs range from 144 to 911. Shift is excluded. T means temperature. Read correlated factors together; correlation does not establish causal direction.
09
MODEL DESIGN

XGBoost classifies outlet AO

The target is the outlet AO measurement matched to each input record.

HIGH AO
Y > μtrain

Above the training mean

LOW AO
Y ≤ μtrain

At or below the training mean

Random modelμtrain = 79.059473Chronological validation modelμtrain = 78.990926

XGBoost returns p(high AO). The predicted class is high when p ≥ 0.5. Thresholds, category mappings and quartiles come from training data only.

The classifier targets AO class. Continuous decomposition efficiency is computed separately from the measured chemistry. Training uses 60 candidate configurations and five-fold cross-validation.
10
EXPLANATION DESIGN

Two explanation branches

Compare the predicted class with the class of the corresponding measured outlet AO.

Predicted and measured AO classes match

Example: predicted high AO, measured high AO.

  1. Select factors whose local SHAP supports the measured class.
  2. Double their weights for values outside Q25–Q75.
  3. Convert weights to percentages and rank the factors.

Local SHAP explanation

Predicted and measured AO classes differ

Example: predicted low AO, measured high AO.

  1. Use the measured AO class: high or low.
  2. Select factors outside Q25–Q75 whose historical rules support that class.
  3. Rank them using historical SHAP support and calculate weights totaling 100%.

No matching rule: reason unknown.

Q25–Q75 is the middle 50% of training values. Actual outlet AO is required for this diagnosis. Rule explanations leave the original prediction unchanged.
11
WORKED EXAMPLE 1

Local SHAP with quartile-aware ranking

Example: Jun 24, 2026 Measured AO 82.8Predicted and observed class: high
Supporting factorPositive SHAPMultiplierFinal weight
Middle temperature 62.12565 > Q750.406441× 265.84%
Inlet NK 138.5 within Q25–Q750.266847× 121.61%
Seed solids 624.3 < Q250.053380× 28.65%
Inlet AO 151.2 within Q25–Q750.048083× 13.89%
Weighti = mi × positive SHAPiΣ mj × positive SHAPj × 100%
Middle-temperature support rises from 52.46% to 65.84%. Values outside Q25–Q75 get multiplier 2; equality stays inside. Weights represent relative support rather than physical loss allocation.
12
WORKED EXAMPLE 2

Directional rules for a model mismatch

Example: Jan 23, 2026 Measured AO 81.3Predicted low, observed high
Matched high-AO ruleGlobal positive scoreFinal weight
Seed solids 686.5 < Q25 = 699.30.05592250268.39%
Inlet AO 152.6 > Q75 = 152.10.02584439631.61%

Learn the direction

A tail supports a class when its mean signed training SHAP is positive, above tolerance.

Calculate the weight

Score = mean positive class-SHAP across all training rows. Normalize across matched factors.

The result offers historical candidate drivers when local prediction disagrees.

No extra multiplier applies. Rules can disagree with local SHAP. Missing, within-range and opposite-direction inputs do not match; shift has no numeric tail rule.
13
EVALUATION DESIGN

Two training and test designs

Random model

911 training records228 test

Random 80% / 20% split with a fixed seed.

Chronological validation model

OctNovDecJanFebMarAprMayJunJulAugSepOct*
Oct 2025–Aug 2026
994 training recordsSep–Oct 19, 2026
145 test records

The same model and explanation design is refitted using the earlier 11 months.

*October is incomplete. Historical chronological evaluation has no 48-hour label purge: 6 training labels reach or cross the test boundary. Inner selection remains random five-fold CV.
14
RESULTS · RANDOM MODEL

Random-split explanation coverage

95.65%

of the lowest-efficiency
10% and 20% test cohorts
receive supporting factors.

Test cohortNSHAPRulesUnknownCoverage
All test records22817547697.37%222/228
Lowest efficiency 10%23211195.65%22/23
Lowest efficiency 20%46359295.65%44/46
Cohorts use the lowest measured efficiencies. Coverage counts records with SHAP or rule support and is separate from prediction accuracy or verified causal correctness.
15
RESULTS · CHRONOLOGICAL VALIDATION

Chronological explanation coverage

100%

of the lowest-efficiency
10% and 20% historical test cohorts
receive supporting factors.

Test cohortNSHAPRulesUnknownCoverage
All test records1457764497.24%141/145
Lowest efficiency 10%151320100.00%15/15
Lowest efficiency 20%292090100.00%29/29
15/15 and 29/29 are historical cohort results, not a guarantee for future data. Combined coverage includes both model support and historical rule support.
16
CONCLUSIONS

Project summary

  • Separate monitoring and factor-explanation modules support abnormality detection in Bayer-process precipitation.
  • Twelve process factors combine work shift, tank temperatures and liquor chemistry from 1,139 historical records.
  • XGBoost and local SHAP work with historical directional rules to rank candidate drivers of abnormal outcomes.
  • Explanation coverage reaches 95.65% in the lowest-efficiency 10% and 20% random-test cohorts, and 100% in the corresponding chronological cohorts of 15 and 29 records.
Coverage describes availability of supporting factors for the observed outcome. Causal confirmation requires additional engineering evidence.
17

Slides