Process Twin — Reference Architecture

How HJT shop-floor signals flow through Google Cloud, Vertex AI, Gemini and MDE into physics-grounded intelligence and back to the operator.

Architecture at a GlanceSix logical layers — from cell/module signals to closed-loop action

Source Systems

PLC/SCADA, MES, EL/AOI vision, LIMS, ERP

Edge & Ingestion

OPC-UA, MQTT, MDE gateways, Pub/Sub

GCP Data Fabric

BigQuery, Cloud Storage, Dataplex, Feature Store

Physics + Vertex AI

First-principles twin · Vertex models · Gemini

Applications

Process Twin cockpits · EyeZ · Ved copilot

Action & Loop

Setpoint write-back · CMMS · notifications

Vendor-neutral · brownfield-ready · designed for RIL's 20 GW HJT reference line

01 · Source SystemsEvery operational signal the HJT Process Twin learns from

OT · Cell & Module Line

PLC / DCS

Texturing, PECVD, TCO sputter, printing, stringer, laminator PLCs

SCADA & Recipes

Setpoints, alarms, recipe books per station

Process sensors

RF power, chamber pressure, plate T, vacuum, KOH bath

IIoT & vibration

Motor health, pump vibration, energy sub-meters

Operations

MES

Wafer / cell / module traceability, WIP, lot genealogy

Historian

OSIsoft PI / AVEVA — multi-year time series at 1s

LIMS / QC

Flash-test, EL grading, incoming wafer specs

CMMS / EHS

SAP PM — PM schedules, work orders, incidents

Business

SAP S/4HANA

Material, cost, inventory, bin-to-line consumption

SCM / Dispatch

Module shipment, RMA, warranty registry

QMS / SOPs

Change control, spec revisions, audit trails

HRIS / Shift

Crew rosters, operator certifications

Vision & External

EL & AOI

String EL, post-lam EL, glass/frame AOI cameras

In-line vision

Print inspection, ribbon alignment, bubble scans

Market feeds

Poly / silver / EVA indices, module price benchmarks

Weather / grid

Ambient T/RH — feeds lamination & TCO drift models

Read-only by default · no rip-and-replace of existing OT/IT

data flow

02 · Edge & Ingestion — Google MDEManufacturing Data Engine (MDE) as the OT↔cloud bridge — secure, normalized, streaming

Google MDE EdgeMDE

On-prem appliance · OPC-UA / Modbus / MQTT connectors · store-and-forward on link loss

MDE Cloud IngestMDE

UNS-style normalized tags · schema registry · lands directly into BigQuery + Pub/Sub

Cloud Pub/SubGCP

Streaming backbone · 50–200k tags @ 1s · fan-out to twin, analytics, alerts

Vision PipelineVertex

RTSP → GKE inference (Vertex-served ONNX/TF) · EL/AOI defect metadata to Pub/Sub

Batch ConnectorsGCP

Datastream + Cloud Storage · SAP / LIMS / historians via SFTP/OData/IDoc

OT Security LayerGCP

Read-only default · TLS 1.3 · VPC-SC perimeter · IAM + workload identity

MDE Edge on shop-floor appliances · one-way diode optional · ISA/IEC 62443-aligned

data flow

03 · GCP Data FabricSingle source of truth for physics, process and business — on Google Cloud

BigQueryGCP

Partitioned + clustered time-series · hot 90d / warm 5y · sub-second cockpit queries

Cloud Storage — Bronze/Silver/GoldGCS

Raw → conformed → KPI marts · Dataflow / Dataform pipelines

Asset & Genealogy GraphSpanner

ISA-95 hierarchy · wafer → cell → string → module lineage

Semantic / KPI LayerLooker

Looker semantic model · η, FF, Pmax, FPY, OEE, kWh/module, ₹/W

Dataplex GovernanceGCP

Catalog · field-level lineage · DLP masking · audit log

Vertex Feature Store + Vector SearchVertex

Online features for surrogates · embeddings for Ved RAG on SOPs / RCA history

Open formats (Parquet / Iceberg) · Dataplex catalog & lineage · India data residency

data flow

04 · Process Twin Intelligence — Physics + Vertex AI + GeminiFirst-principles digital twin fused with Vertex AI models and Gemini reasoning

Physics Twin Core

PECVD, TCO sputter, print/cure, stringer, Kamal-Sourour lamination — first-principles ODE/PDE surrogates

VedAI ModelsVertex

Yield / drift / RUL / defect models · custom training + AutoML · MLOps via Vertex Pipelines

Ved ReasoningGemini

Gemini 2.x on Vertex — Ved reasoning · RCA drafting · NL-to-SQL over BigQuery

Forecast & SPC

P10/P50/P90 cones on η, Pmax, FPY · Nelson-rule SPC on 100+ tags

Bayesian Root Cause

Causal DAGs · 'why did bubble rate spike?' answered with weighted drivers

Anomaly & Drift

Multivariate drift on assays, RF power, plate T, vacuum, vibration

Scenario Simulator

What-if lever swings · sensitivity tornado · cost/watt trade-offs

EyeZ Vision ModelsVertex

EL microcrack / dark-cell / bubble classifiers · served via Vertex Endpoints

Guardrails & Approvals

Operator-in-loop · safe action envelopes · immutable approval log

Ved Model RegistryGemini

Ved Model Registry · Vertex Pipelines · drift monitors · shadow-mode rollouts

Trust layer

Physics-grounded confidence · SHAP drivers · counterfactual outcome per recommendation

Human gates

Role-based approvals · every write-back logged · reversible by design

Every recommendation carries physics rationale, confidence, drivers and counterfactual

data flow

05 · Applications & User ExperienceOne fabric, many surfaces — cockpits, assistant, and closed-loop workflows

Process Twin Cockpits

Overview · line topology · per-station deep physics

Process Analytics

SPC, drift, sensitivity tornado, correlation drill-down

EyeZ — Vision QC

Final EL gallery · quality inspection · defect reference library

Defect Genealogy

Wafer→cell→string→module provenance · blast-radius clusters

Ved AssistantGemini

Page-aware NL assistant · grounded on Twin + BigQuery + SOP vectors

Scenario Simulator

Physics-lab what-ifs before commit · recipe A/B

Model Registry

Every deployed model — version, owner, drift, business impact

Reports & ESG

Shift reports, audit packs, kWh/module, carbon disclosures

Same BigQuery + Vertex source of truth · no copies · consistent numbers across every screen

closed loop

06 · Action & Integration LayerInsights become actions — with humans firmly in the loop

Setpoint Write-Back

Approved recipe adjustments to PLC/SCADA via MDE reverse channel

CMMS Work Orders

Auto-drafted SAP PM notifications from RUL / anomaly triggers

MES Disposition

Cell / module hold, rework, scrap decisions written to MES

Operator Notifications

Severity-routed alerts · mobile push · shift-handover packs

Every write-back is approval-gated, audited, and reversible

Cross-cutting · Security & ComplianceDesigned to pass RIL group-CISO and plant-IT review

ISA/IEC 62443 OT segregation · optional unidirectional gateway on MDE
GCP: VPC-SC perimeter · CMEK · Assured Workloads · SOC 2 / ISO 27001
SSO via Reliance IdP · IAM roles per persona (operator / engineer / plant head)
Immutable audit log — every recommendation, approval and setpoint write
India data residency — Mumbai / Delhi GCP regions

Cross-cutting · Deployment & ScaleFrom pilot line to 20 GW rollout

Hybrid: MDE edge on-prem + GCP control plane (Mumbai region)
GKE-native services · horizontally scalable per plant / per line
BigQuery + Pub/Sub backbone · 200k+ tags @ 1s · sub-second cockpit refresh
Multi-tenant by plant · shared Vertex models, isolated BigQuery datasets
12-week pilot on one HJT line → full plant in 6 months → 20 GW rollout