Skip to content
Research

Making manufacturing processes predictable before they are proven

Nine years chasing process failures downward — from the toolpath to the microstructure — and the argument for where that leads next.

The argument

Why this problem, and why this method

Seven movements, in order. Each follows from the one before it rather than sitting beside it.

  1. R1

    Why manufacturing

    I did not arrive at manufacturing through reading about it. I arrived at a machine in 2014 with a toolpath to write, where a wrong decision showed up in the surface finish within minutes. Manufacturing is where physics stops being theoretical: a tolerance either holds or it does not, and the part in your hand is the argument. Nine years later that is still what holds my attention — not the theory of a process, but the gap between what a process is supposed to do and what it actually does on a Tuesday afternoon.

  2. R2Published — Peer-reviewed output exists

    Why materials

    Chasing that gap pushed me downward. A rejected part is a process event; a process event is usually a materials event wearing a process costume. Tooling taught me that geometry is negotiable and material behaviour is not. That is what took me to NTUST, where I stopped treating material as a given and started treating it as the variable — designing an infiltrated stir casting route for a magnesium hybrid metal composite, and learning that the property claim is worthless until the microstructure explains it.

  3. R3Proposed — Stated research intent; no work done yet

    Why additive manufacturing

    In casting, microstructure is largely inherited. In laser powder bed fusion it is designed — every scan vector contributes to the thermal history that decides the local microstructure, which means the process parameters are material parameters. That is exactly why it is hard: the design space is enormous, the behaviour is path-dependent, and qualification still runs on build-and-inspect. It is the process family where the questions I have been chasing since 2014 concentrate most sharply. I want to work on it; I have not worked on it yet.

  4. R4Proposed — Stated research intent; no work done yet

    Why physics-informed machine learning

    This is the methodological argument, and it is the reason the other three connect. Designed experiments — Taguchi, factorials, ANOVA — are the instruments I know best, and their limits are structural: they screen a modest factor count against an assumed response surface. L-PBF presents dozens of coupled factors and a thermal history no polynomial represents honestly. The obvious alternative fails for the opposite reason: purely data-driven models need data volumes that an expensive characterisation campaign cannot produce, and outside the sampled region they extrapolate to answers that violate conservation laws. Embedding the physics in the model — as a constraint, as a residual term, as a reduced-order prior — is how you get predictive power from small, costly datasets. Manufacturing research only ever has small, costly datasets. That is the direction I want my doctoral work to take.

  5. R5

    Future vision

    Processes qualified by prediction rather than by exhaustive trial. Today, introducing a new material-process combination means paying for a campaign of builds and inspections to map a window empirically, and paying again for the next material. If the physics is in the model, most of that cost becomes computation and the experiments become confirmatory rather than exploratory. That shift is what would make advanced manufacturing accessible to organisations that cannot fund a hundred trial builds.

R6 · Potential research

Threads, labelled by maturity

One published, three proposed. A proposed thread states what starting it would require, because ambition with a stated cost is the only kind worth reading.

Published — Peer-reviewed output exists

Magnesium hybrid metal composites

Can infiltrated stir casting distribute hybrid reinforcement uniformly enough to raise specific energy absorption without losing castability?

Lightweight crash structures are judged on energy absorbed per unit mass. Magnesium starts from the best possible density and the worst possible stiffness, so the question is whether reinforcement can be added by a route a foundry could reproduce.

  • Infiltrated stir casting
  • Taguchi method
  • ANOVA
  • SEM / EDX
  • XRD
  • Mechanical testing
Proposed — Stated research intent; no work done yet

Process quality at manufacturing scale

At gigafactory volume, which defect modes are genuinely explainable from routine line data, and which require characterisation that production cannot afford to run?

High-volume battery manufacturing produces more process data than any team can interpret, and the cheapest measurements are rarely the ones that carry the causal signal. Knowing which is which decides where inspection ends and process control begins.

  • Statistical process control
  • Root cause analysis
  • Defect mode classification

No work done yet — this would require

  • Clearance to publish anonymised or synthetic process data
  • Characterisation access alongside production sampling
  • A defined defect taxonomy that survives across cell formats
Proposed — Stated research intent; no work done yet

Metal additive manufacturing — L-PBF process qualification

Can laser powder bed fusion process windows be qualified predictively, instead of by exhaustive build-and-inspect trials?

L-PBF makes microstructure a designed variable rather than an inherited one — and simultaneously makes it path-dependent, which is why qualification currently depends on post-hoc CT inspection and campaign after campaign of trial builds.

  • Design of experiments
  • Metallography
  • Non-destructive evaluation

No work done yet — this would require

  • An L-PBF platform with in-situ melt-pool monitoring
  • CT and metallographic ground truth for defect labelling
  • A labelled dataset spanning more than one material system
Proposed — Stated research intent; no work done yet

Physics-informed machine learning for manufacturing processes

Can known process physics be embedded in a learned model well enough to give predictive power from the small, expensive datasets a real experimental campaign produces?

This is the methodological question behind the other three. Classical designed experiments assume a response surface and handle a handful of factors. Purely data-driven models need volumes of data that powder-bed or production characterisation campaigns cannot produce, and they extrapolate unphysically outside the sampled region. Constraining the model with conservation laws and known process behaviour is the route to prediction from small data — which is the only kind of data manufacturing research actually has.

  • Reduced-order process modelling
  • Design of experiments
  • Statistical validation

No work done yet — this would require

  • A labelled process dataset with physical ground truth
  • A reduced-order process model to supply the physics constraint
  • Computational resources for training and validation
  1. R7

    Why this matters

    Battery and semiconductor manufacturing are scaling faster than process understanding. I see one of those ramps every working day — high-volume lines where a deviation is measured in parts per million, where the data volume is enormous and the explanation is thin. The gap between what a plant measures and what it understands is not an academic curiosity there; it is counted in production days. That is the gap I want to spend a doctorate on, and it is the one place where my nine years on the floor is not a detour from research but the reason I know which question to ask.

Methods

What I can run, and where I ran it

Every method links to the work that evidences it. A method with no attached result is a vocabulary claim.

Methods and equipment, with where each was applied and the work that evidences it
MethodTypeWhere appliedEvidence
Infiltrated stir castingTechniqueFabrication route for the magnesium hybrid metal compositeCase study
Taguchi methodStatisticalFactor screening across the composite casting campaignCase study
ANOVAStatisticalEstablishing which process factors were statistically significantPublication
Design of experimentsStatisticalExperimental campaigns and process trials—
SEM / EDXEquipmentReinforcement distribution and elemental mappingPublication
XRDEquipmentPhase identification in the compositePublication
Mechanical testingTechniqueEnergy absorption and mechanical response of the sample setPublication
Statistical process controlStatisticalBattery cell production quality at CATLExperience
Root cause analysisTechniqueDefect investigation on high-volume productionExperience
Metrology & first-article inspectionTechniqueTooling and machined component verification—

Software

  • SolidWorks

    Machine and fixture design

  • Siemens NX

    Modelling and CAM

  • AutoCAD

    Fabrication drawings

  • Inventor

    Industrial machinery design

  • PowerMill

    Multi-axis CNC programming

  • TypeScript / React / Next.js

    This site, designed and built end to end

Publications

Peer-reviewed output

One article. Full metadata, author position, and citation export — specificity substitutes for volume.

Published — Peer-reviewed output existsPeer-reviewed

Magnesium Hybrid Metal Composite via Infiltrated Stir Casting Technique: A Novel Approach to Enhance Energy Absorption

Journal of Applied Science and Engineering · Vol 29(2) · 2026

Authors

M. Munajad (author 1)

The complete author list as printed is pending confirmation and will be published here in full. Author position is not being represented as more than it is.

This work introduces an infiltrated stir casting technique to fabricate a magnesium hybrid metal composite, targeting improved energy absorption for lightweight structural applications. Microstructure was characterized with SEM/EDX and XRD, and mechanical performance was evaluated to validate the novel processing route.

Magnesium CompositesStir CastingEnergy AbsorptionSEM/EDXXRD
View on publisherDOI: 10.6180/jase.202602_29(2).0004

BibTeX export is disabled until the complete author list is confirmed. Exporting incomplete metadata would propagate errors into other people's reference managers.

Roadmap

Three horizons

Where the work goes over the next decade, and which part of it I am seeking supervision for.

NowActive — Work in progress; not yet published

0 – 1 year

Consolidate method at production scale

Deepen the process-quality work at CATL and formalise the methods I use daily — defect taxonomy, control-plan reasoning, and where routine line data stops being able to explain a failure.

DoctoralProposed — Stated research intent; no work done yet

1 – 4 years

Physics-informed models for process qualification

Build and validate physics-constrained surrogate models for a metal additive process, against experimental ground truth. This is the work I am seeking supervision for.

BeyondProposed — Stated research intent; no work done yet

4+ years

Predictive process qualification as practice

Move the approach from a demonstrated method to something a manufacturing organisation can adopt — which means transferable models, stated uncertainty, and qualification evidence a regulator or a customer will accept.

Laboratory fit

What I am looking for in a group

Stated as capabilities rather than institutions — a page claiming deep alignment with every target lab would be transparently generic.

I am looking for a group where experimental process work and data-driven modelling sit in the same room rather than in different buildings. Concretely, the capabilities that would make the work possible:

Laboratory capabilities sought

  • A metal additive platform — L-PBF preferred — with in-situ process monitoring
  • High-speed thermal imaging or equivalent melt-pool instrumentation
  • Metallography, SEM/EDX, and EBSD access for microstructural ground truth
  • X-ray CT for internal defect characterisation
  • Computational resources for model training and validation

Supervision shape sought

  • Co-supervision spanning manufacturing process engineering and data-driven modelling
  • A group that treats industrial data access as an asset rather than a complication
  • Willingness to publish negative and boundary results, not only successful process windows
Collaboration

Three shapes this could take

Doctoral supervision

I am seeking a PhD position in advanced manufacturing, metal additive manufacturing, or physics-informed modelling of manufacturing processes. My research question is being finalised with prospective supervisors rather than declared in advance.

Industrial data access

I work inside high-volume battery manufacturing. Where confidentiality permits, I can help a research group frame questions against how production processes actually behave rather than how they are assumed to behave.

Method transfer

Designed experiments, statistical process control, and materials characterisation applied to a partner's process problem — the practical end of the same method.

Profiles & provenance

Where to verify the record

Provenance rather than impact metrics. Citation counts are not shown while they would be near zero — the number would draw attention to the weakest available comparison.

Venue
Journal of Applied Science and Engineering
Volume / issue
29 (2)
Year
2026
DOI
10.6180/jase.202602_29(2).0004
Peer reviewed
Yes

ORCID, Google Scholar, and a downloadable academic CV will be linked here. They are omitted until they exist rather than shown as placeholders.

For research groups

Considering a doctoral candidate with nine years on the floor?

I am looking for a group where experimental process work and data-driven modelling sit in the same room. Happy to discuss fit before either of us commits to an application.

Discuss supervision or collaboration