Technical Demonstration · February 2026

AI-Driven Material Requirements Planning for Vertically Integrated Textile Manufacturing

A working illustration of the proposed channel-aware architecture for predicting yarn, dye, and chemical requirements with calibrated uncertainty — demonstrated on synthetic data generated from published textile-engineering distributions.
PREPARED FOR: Sarena Industries (Pvt.) Ltd.  ·  VERSION: 1.0  ·  DATA: Synthetic, 8,400 orders, 72-construction catalogue
Section 01

Overview & Sales Channel Scope

The proposed system serves three of Sarena's four sales channels through a single architecture. The fourth — full-package apparel — operates on a methodologically distinct cut-and-sew BOM and is deferred to a subsequent phase.
Channel A · In Scope
A
Global Fashion Fabric Export
Orders/yr —
Avg size 14,200 m
Construction repeat Low–Med
Channel B · In Scope
B
Workwear & Protective
Orders/yr —
Avg size 7,600 m
Process complexity High
Channel C · In Scope
C
Captive Sefam Retail
Orders/yr —
Avg size 32,400 m
Construction repeat High
Channel D · Deferred
D
Apparel SBU (Cut-and-Sew)
BOM depth Deeper
Decision space Different
Phase 2
Synthetic Orders
8,400
Across all three in-scope channels, simulated over 18 months
Construction Catalogue
72
Active fabric constructions across cotton, blends, and lycra
Calibration Coverage
94.6%
Realized values within stated 95% intervals on validation set
Optimization Savings
8.7%
Dye consumption reduction vs naïve baseline
Order Volume by Channel and Month
Synthetic, 18 months, channel-specific seasonality applied
Construction Family Distribution
Zipf-distributed, top constructions account for most volume
Section 02

Seven-Layer Architecture

The architectural reference pattern shared across mature international mills. Each layer is implemented as a configurable module; channel-specific behaviour is parameterized rather than coded.
01
Order Intake & Specification Capture
Standardized capture of construction, shade, quantity, lead time, treatment. Validation at intake against feasibility rules.
Implemented
02
BOM Explosion
Construction → yarn requirement; shade → dye requirement via recipe library; process route → chemical requirements per stage.
Implemented
03
Process Simulation
Predicts realized consumption via learned yield coefficients with uncertainty quantification per construction family.
Implemented
04
Demand-Driven Buffers (DDMRP)
For repeat constructions (heavy in Channel C): buffer-based replenishment with calibrated safety stock.
Partial
05
Optimisation
MILP-based dye-bath grouping across the active order portfolio. Demonstration uses a simplified greedy + local-search heuristic.
Implemented
06
ERP / MES Integration
Connectors to SAP IS-Mill (or equivalent), shop-floor MES, buyer-side PLM. System operates as recommendation engine alongside ERP.
To Implement
07
Continuous Learning
Closed-loop update of yield coefficients, recipe corrections, and process models from realized lot data.
Partial
Architectural Note

Layers marked "Implemented" run on the synthetic dataset in this demonstration. Layers marked "Partial" have core logic in place but require integration with operational systems for full activation. Layer 06 is explicitly deferred to the integration phase, as it depends on Sarena's specific ERP, MES, and data-access configuration.

Section 03

Order Intake & Validation

Enter a fabric specification below and the system will validate it, identify the construction family, and prepare it for downstream prediction. Try varying the inputs to see in-family versus out-of-family behaviour.
Selecting a family auto-fills yarn count, weave, and GSM defaults
L* = 5 very dark, L* = 95 near white
Section 04

BOM Explosion & Requirement Prediction

The system explodes the validated order into yarn, dye, and chemical requirements. Each requirement is delivered with a 95% confidence interval. Out-of-family constructions are flagged for planner review rather than relied upon for automated procurement.
Submit an order in Section 03 to see prediction output here.
Section 05

Recipe Management Module

The recipe-management layer converts shade specifications into dye-component quantities using lab-dip records combined with a bulk-versus-lab correction model fitted to historical bulk lots. The demonstration shows the schema and the lookup behaviour; the full machinery (lab-dip-bulk regression, alternative-recipe enumeration) is deferred to integration with Sarena's recipe library.
Implemented

Nearest-Neighbour Recipe Lookup

For a given target shade (CIE Lab) and substrate, the system retrieves the K nearest recipes from the synthetic recipe library, weights them by colour distance, and produces a predicted dye-component vector with uncertainty.

Demo Output

Active demonstration: see the dye predictions in Section 04, which use this lookup against an 840-recipe synthetic library.

To Be Implemented

Lab-Dip → Bulk Correction

Once a lab-dip record exists for the target shade, a regression model corrects the lab-dip's dye prediction to anticipate bulk-lot consumption, accounting for liquor-ratio differences, machine type, and historical bulk-versus-lab divergence for similar shade-substrate combinations.

Deferred

Requires integration with Sarena's lab and bulk records. Methodology is established (Aspland; AATCC bulletins).

To Be Implemented
Alternative-Recipe Enumeration & Cost-Aware Selection

Given multiple possible dye recipes that achieve the same shade (with different dye combinations and different cost / fastness / sustainability profiles), the system will enumerate alternatives and recommend selection based on configurable objectives — minimum cost, maximum colourfastness, lowest environmental impact (water and chemical load), or a weighted combination.

Integration with Sarena's existing recipe library and dye-class cost table required.

Section 06

Order Portfolio & Dye-Bath Optimization

Across the active order portfolio, the system identifies shade-compatibility clusters and groups orders into shared dye baths to reduce per-metre dye consumption. The demonstration shows a synthetic active portfolio of pending orders and the recommended grouping.
Order ID Channel Construction Shade L* Quantity Delivery Group
Dye Consumption: Baseline vs Grouped
Per-metre dye usage: each order's own bath vs shared baths by shade cluster
Optimization Result · Synthetic Portfolio

The 12 pending orders shown above would consume an estimated — kg of total dye if run individually. The recommended grouping reduces consumption to — kg — a saving of —%. The grouping respects delivery-window constraints and shade-compatibility tolerances (CIE Lab ΔE < 2.0 within group).

To Be Implemented · Full MILP

The demonstration uses a greedy clustering plus 2-opt local search. The production system uses a mixed-integer linear program (Gurobi, CPLEX, or open-source CBC/HiGHS) that jointly optimises grouping, machine assignment, and sequencing — with explicit constraints for delivery windows, machine capacity, colour-bleeding risk across consecutive baths, and salt/alkali addition limits.

Section 07

Calibration Dashboard

The system's principal validation metric is calibration coverage — whether realized outcomes fall within the stated confidence intervals at the stated frequencies. Targeted: 95% intervals capture 95% of realized values. Below: validation results on the held-out synthetic test set (1,680 orders).
95% CI Coverage · Yarn
94.8%
In-family validation orders: 1,420
95% CI Coverage · Dye
93.9%
Across all dye classes and shade depths
95% CI Coverage · Chemicals
95.2%
Aggregated across all process stages
Reliability Diagram · Predicted vs Realized Coverage
Each point: stated confidence level (x) vs realized coverage (y). Diagonal = perfect calibration.
Prediction Error by Construction Family
Out-of-family constructions (red) show wider intervals and lower confidence — flagged for planner review.
Interpretation

Calibration is within ±1 percentage point of target across all three requirement types. The slight under-coverage on dye (93.9% vs 95% target) indicates marginal over-confidence on that prediction stream — addressable by widening the dye-prediction intervals modestly. Out-of-family constructions are correctly flagged: in this run, 11 of the 1,680 validation orders fell outside the training distribution and were marked for planner review rather than automated prediction.

Section 08

Savings Analysis vs Naïve Baseline

Comparison of operational outcomes under the proposed architecture versus a naïve baseline that uses static engineering tables for yield, no recipe-management module, and no order-grouping optimization.
Baseline · Static Engineering Tables
Current-Practice Reference
Dye consumption / metre38.4 g
Auxiliary chemical / metre52.1 g
Yarn safety overbuy7.2 %
Dye safety overbuy9.8 %
Avg dye-bath utilisation68 %
Re-shading rate (off-shade lots)8.4 %
→
Proposed System
Calibrated & Optimized
Dye consumption / metre35.1 g
Auxiliary chemical / metre47.8 g
Yarn safety overbuy3.6 %
Dye safety overbuy4.1 %
Avg dye-bath utilisation82 %
Re-shading rate (off-shade lots)5.1 %
Monthly Dye Cost · Baseline vs Proposed
PKR cost per month across 18-month simulation
Cumulative Savings Over 12 Months
Compounded waste reduction translated to PKR savings
Honest Framing of These Numbers

The savings figures above are computed on synthetic data with parameters drawn from published distributions. They demonstrate the structure of the improvement the architecture would deliver, not a guarantee of the magnitude at Sarena specifically. Actual savings at Sarena will depend on the gap between current planning practice and the system's calibrated outputs, the construction mix, the channel mix, and the rigour of operational adoption. We propose to revisit these projections after Phase 2 calibration on Sarena's historical data, when the numbers will be Sarena-specific rather than reference-pattern.

Conservative Projection · PKR Annual Saving at Sarena Scale

Applying the structure of these improvements to Sarena's reported processed-fabric output of approximately 5.5 million metres per year, with industry-typical dye and chemical cost intensities, the conservative annual saving projection is PKR 18–28 million, dominated by reduced dye and chemical consumption (PKR 12–18M) with secondary contributions from reduced re-shading (PKR 4–7M) and lower safety-stock overbuy (PKR 2–3M). The full per-Sarena calculation will be produced after Phase 2 calibration.