Research starter kit: from reading to first contribution
📍 Where we are: Part VI · Starting Your Own Research — Chapter 18 of 19. Chapter 17 showed that a result depends on how you split the data; this chapter shows how to find the literature around such a result, read it critically, and set up a project others can reproduce.
Starting research is like moving to a new city and wanting to know the best bakery. You ask several people, because each knows different streets. You do not simply pick the shop with the longest queue, because the oldest shop has had decades to build a queue and the new one may be better. You taste before you recommend. And you write down the route so a friend can find the same shop. This chapter is that routine for papers: search in several places, treat counts as hints, read for evidence, and leave a trail others can follow.
What this chapter covers
Five practical skills, with every command run on 2026-11-15: finding foundational and recent papers; using citation counts without being fooled; reading a paper against the evidence levels of this book; knowing the venues and communities; and making your own work reproducible. It ends with a project template, ten first-project ideas each tied to a chapter and a lab, and a pointer to the verdict chapter for choosing a problem.
Problem
A newcomer faces tens of thousands of AM papers and no map. Search engines rank by text similarity, not by trust. Counts reward age and fashion. Many results cannot be reproduced because data or code were not shared, and many ML results are inflated by leakage [1]. The aim here is not to read more; it is to read the right 20 papers well and to build something that survives scrutiny.
State of the art
1. Find the literature
Use at least two sources, because each index covers different documents. All three below have free APIs [2][3][4]. This book's helper scripts/cite.mjs wraps Crossref and Semantic Scholar/OpenAlex; examples/research-starter-kit/litcheck.mjs (Node 18 or newer, no key) adds cross-source checks. Google Scholar is useful for manual discovery; we found no official API for it, so we did not query it programmatically and show no Scholar output.
Foundational papers: filter by age, sort by counts. Output of our run, 2026-11-15:
$ node scripts/cite.mjs search "additive manufacturing of metallic components process structure properties" --min 500 --rows 5
8836 2018 DebRoy T Additive manufacturing of metallic components – Process, structure and properties [Progress in Materials Science] doi:10.1016/j.pmatsci.2017.10.001
1088 2016 Körner C Additive manufacturing of metallic components by selective electron beam melting — a revie [International Materials Reviews] doi:10.1080/09506608.2016.1176289
$ node scripts/cite.mjs search "keyhole porosity laser powder bed fusion" --from 2019 --min 200 --rows 5
432 2022 Huang Y Keyhole fluctuation and pore formation mechanisms during laser powder bed fusion additive manufacturing [Nature Communications] doi:10.1038/s41467-022-28694-x
358 2020 Ronneberg T Revealing relationships between porosity, microstructure and mechanical properties of lase [Materials & Design] doi:10.1016/j.matdes.2020.108481
337 2023 Ren Z Machine learning–aided real-time detection of keyhole pore generation in laser powder bed fusion [Science] doi:10.1126/science.add4667
The top hit is a long review of metal AM processes, structures and properties [5]; the keyhole paper that used synchrotron X-ray imaging [6] is a good first paper to practise the checklist below on. cite.mjs search re-sorts the top 150 relevance hits by Crossref count, so a too-strict --min returns nothing. We ran search "laser powder bed fusion review" --min 1000 and it printed no lines at all: silence means "no hit in the pool", not "no such paper". Lower --min or reword.
Recent papers: constrain the date, then judge the venue.
$ node scripts/cite.mjs search "machine learning additive manufacturing" --from 2024 --min 30 --rows 4
124 2025 Faegh M A review on physics-informed machine learning for process-structure-property modeling in a [Journal of Manufacturing Processes] doi:10.1016/j.jmapro.2024.11.066
110 2024 Farrag A Physics-Informed Machine Learning for metal additive manufacturing [Progress in Additive Manufacturing] doi:10.1007/s40964-024-00612-1
106 2025 Inayathullah S Review of machine learning applications in additive manufacturing [Results in Engineering] doi:10.1016/j.rineng.2024.103676
83 2024 Akbari P Machine learning prediction of mechanical properties in metal additive manufacturing [Additive Manufacturing] doi:10.1016/j.addma.2024.104320
Reading a hit before trusting it takes one command (the abstract comes from Semantic Scholar, with OpenAlex as fallback; output trimmed):
$ node scripts/cite.mjs abstract 10.1038/s41467-022-28694-x
# Keyhole fluctuation and pore formation mechanisms during laser powder bed fusion additive manufacturing (2021, Nature Communications) S2 cited-by: 490
Keyhole porosity is a key concern in laser powder-bed fusion (LPBF), potentially impacting component fatigue life. ... Using synchrotron X-ray imaging we reveal keyhole and bubble behaviour ...
Snowball in both directions. Take one strong paper and follow its references (backward) and the works that cite it (forward). OpenAlex does the forward step directly; here are the 2026 papers citing the foundational review above:
$ curl -s "https://api.openalex.org/works?filter=cites:W2762367303,from_publication_date:2026-01-01&sort=cited_by_count:desc&per-page=3&select=doi,title,publication_year,cited_by_count"
citing works in 2026 so far: 1696
https://doi.org/10.1007/s43452-026-01520-7 | 2026 | 71 | Microstructure and mechanical properties of Ni superalloy fabricated b...
https://doi.org/10.1016/j.jmrt.2026.01.173 | 2026 | 42 | Advancements in laser processes for additive Manufacturing: Techniques...
https://doi.org/10.1016/j.jallcom.2026.186347 | 2026 | 34 | Wire Arc Additive Manufacturing (WAAM): A comprehensive review of proc...
(The W2762367303 id came from the OpenAlex record of that paper's DOI; we piped the JSON through a short Python line to print the table.) A free-text search is fuzzier. Our OpenAlex query search=AM+Bench+melt+pool, sorted by citations, returned a photopolymer resin paper as its top hit. Always read titles.
2. Citation counts, and how they mislead
Same four papers, three databases (node examples/research-starter-kit/litcheck.mjs counts ..., run 2026-11-15):
doi | Crossref year,cites | OpenAlex year,cites | SemanticScholar year,cites | OpenAlex cites/yr
10.1016/j.pmatsci.2017.10.001 | 2018, 8836 | 2017, 9214 | 2018, 8066 | 921
10.1038/s41467-022-28694-x | 2022, 432 | 2022, 528 | 2021, 490 | 106
10.1126/science.add4667 | 2023, 337 | 2023, 342 | 2023, 311 | 86
10.1007/s40192-024-00372-4 | 2024, 15 | 2024, 17 | 2024, 15 | 6
A research start in five steps, and the reminder that one paper has different citation counts in different databases.
Original diagram by the authors, created with AI assistance.
Pitfalls these numbers show or imply:
- Databases disagree. The first paper has 8,836, 9,214 or 8,066 citations depending on the source, and different publication years (2018 in Crossref and Semantic Scholar, 2017 in OpenAlex, probably an online-first date; we did not check). Quote the source and the date with any count.
- Age bias. A count is cumulative, so older papers win. The last paper (2024) has 15 to 17 citations in total, and is cited about 6 times a year, which is not "unimportant"; it is young. Divide by age for a rough rate, and compare only with papers of similar age and field. Field- and year-normalised indicators exist; Waltman reviews how they are built and why the choices matter [7].
- Self-citation. A paper's authors also cite it. We counted, with OpenAlex author ids, how many citing works share at least one author with the cited paper (
litcheck.mjs selfcite): 152 of 9,214 (1.6%) for the review, 24 of 342 (7.0%) for the 2023 Science paper, 40 of 528 (7.6%) for the Nature Communications paper. It is only an approximation, because OpenAlex had no author id for one of the ten authors of the Science paper. An older study of Norwegian papers reported that author self-citations were a much larger share of citations (36%, in the summary we read, not in the full text) [8]. Self-citation is normal; look for it, do not penalise it blindly. - A count says "read by many", not "correct". Reviews collect many citations; a null result may collect few. Use counts to find the field's anchor papers, not to rank the truth.
- Ranking by relevance is not ranking by trust. See the photopolymer hit above.
3. Read a paper against the evidence levels
The preface of this book labels claims demonstrated at lab scale, demonstrated in production or contested. Apply the same labels to every paper you read. Eight questions, in order, take about 30 minutes for a paper you intend to rely on:
- What is the one-sentence claim? Write it down in your own words.
- What was measured, and on what? Alloy, machine, geometry, number of builds or specimens.
- What is the unit of replication? One build? One track? Ten specimens from one build count as one build for most purposes.
- How was the model or comparison validated? For ML, ask what the split respected: build, part, track, or random rows. Lab 5 showed a factor of about 18 between the two [1][9].
- What is the uncertainty, and the noise floor? Is the claimed improvement larger than the repeat scatter?
- Is there an independent check? A second machine or lab, or a blind benchmark such as AM Bench [10].
- Are data and code available, and do they run? A DOI for the data is the best sign [11].
- Which evidence label does it earn? If you cannot place it, say so, as this book does.
4. Venues and communities (verified names)
| Kind | Name | What to know |
|---|---|---|
| Benchmark | NIST AM Bench | NIST-led series of measurements, blind challenge problems and conferences; rounds in 2018, 2022 and 2025; data are archived publicly [12] |
| Standards | ASTM Committee F42 | Additive manufacturing technologies; formed in 2009; subcommittees include Test Methods, Design, Materials and Processes, Data and Terminology [13] |
| Standards | ISO/TC 261 | Additive manufacturing; scope covers processes, terms, process chains, test procedures, quality parameters and supply agreements; ISO and ASTM signed a cooperation agreement in 2011 [14] |
| Conference | Solid Freeform Fabrication Symposium | Annual, University of Texas at Austin; the 36th was in August 2025 [15] |
| Conference | ASTM International Conference on Advanced Manufacturing (ICAM) | Hosted by ASTM's AM Center of Excellence; October 2025 in Las Vegas; focus on standards and qualification [16] |
| Journals (seen in this book's sources) | Additive Manufacturing; Additive Manufacturing Letters; Progress in Additive Manufacturing; Integrating Materials and Manufacturing Innovation; Journal of Manufacturing Processes; Journal of Materials Processing Technology; Progress in Materials Science; Nature Communications; Science | Check a journal's scope and a sample of its papers before submitting; this list is not a ranking |
| Data | NIST Public Data Repository; Materials Data Facility | Citable datasets with DOIs [17][18] |
These are the communities and venues we could verify in this session; others exist. Conference proceedings and standards committees are where methods and test procedures get agreed; journals are where results get argued.
5. Reproducibility practice
- Pin the environment. A
requirements.txtwith exact versions, or a container. Lab 5 pins six libraries and records the Python version. - Verify the data. Download by script, record the DOI and access date, and check a SHA-256. If the source publishes none, pin your own, as Lab 5 does.
- Seed and test. Fix random seeds; add a test that runs twice and compares output, as Lab 5's
make testdoes. - Split by group, and say which group. Name the group key in the README and the paper [9].
- Write a data card. Source, licence, what a row is, known gaps. FAIR principles (findable, accessible, interoperable, reusable) are the general frame [19].
- Containerise the minimum. The template below builds with
docker build.
6. A project template
examples/research-starter-kit/template/ holds a skeleton: README.md, CLAIMS.md (claim, evidence label, data and group key, what would make you retract it), DATA_CARD.md, a pinned requirements.txt, a Dockerfile, a seeded src/run.py that writes the Python and library versions beside the results, and a smoke test. We ran it:
$ cd examples/research-starter-kit && make new NAME=/tmp/my-study
$ cd /tmp/my-study && python src/run.py && python tests/test_smoke.py
{ "python": "3.14.6", "numpy": "2.5.4", "pandas": "3.1.1", "seed": 1, "example_draw": 0.345584192064786 }
smoke test passed
$ docker build -q -t amguide-starter-template . && docker run --rm amguide-starter-template
{ "python": "3.14.10", "numpy": "2.5.4", "pandas": "3.1.1", "seed": 1, "example_draw": 0.345584192064786 }
The first Docker build took 2 min 6 s (downloading the base image and packages) and a rebuild 1.9 s. The container's Python patch version (3.14.10) differs from the host's (3.14.6), yet the seeded value is identical, and the recorded versions show you which is which. Our first container failed with exec: "make": executable file not found: the slim Python image has no make, so the Dockerfile calls python src/run.py directly.
Ten first projects
Each project is an idea, not a result: we did not run them. Each names the chapter it comes from and the lab whose tools it uses. "Lab 1-3" are in Book 3; Lab 4 is Chapter 16; Lab 5 is Chapter 17. Dataset names below are titles we saw in the NIST catalogue; check size, licence and format before you plan around them.
| # | Chapter | Lab | Question | First deliverable |
|---|---|---|---|---|
| 1 | 1 Melt pool | Lab 5 | Does a power law for width hold, and where does depth leave the power-law regime (keyhole)? Add NIST's laser-coupling data, which uses the same case numbers (record mds2-3842) | Width model with a leave-one-case-out table; a plot of residuals against energy per length |
| 2 | 4 Simulation | Lab 4 + Lab 5 | How far are your simulated single-track depths from the 7 measured cases, within the measurement scatter? | A table of simulated vs measured depth and width with the 2.1 µm repeat scatter drawn in |
| 3 | 5 ML | Lab 5 | Pick one published AM-ML paper with shared data; re-run its split by build or part. Does the claim survive? | A one-page replication note with both scores |
| 4 | 6 In-situ monitoring | Lab 5 | On a NIST process-monitoring dataset (the AMMT series), can a layer-wise anomaly baseline beat "predict the mean" when the split is by part? | Baseline with grouped split; check dataset size first |
| 5 | 7 Closed-loop control | Lab 1 | On a virtual printer, how does a controller (e.g. heater regulation) behave under a disturbance, and what does the firmware log show? | A step-response plot from the simulated printer |
| 6 | 2 Defects and residual stress | Lab 3 | How does build orientation change the overhang area and the stress hotspots of the Lab 3 bracket? | Three orientations compared with the printability check and FEA |
| 7 | 12 Qualification | Lab 5 | How many repeats are enough? From the 6 repeats per case, how wide is the interval on a case mean, and what does that mean for a qualification-style claim? | An uncertainty budget table |
| 8 | 13 Digital thread | Lab 5 | Write a data card and a machine-readable metadata file for the Lab 5 table; score it against the FAIR principles [19] | DATA_CARD.md plus a validator script |
| 9 | 15 Cost and throughput | Lab 2 | From a sliced G-code, estimate time and material use per part and how sensitive cost is to layer height | A parametric cost sheet; label every assumption as illustrative |
| 10 | 14 Sustainability | Lab 2 | Estimate the energy of a print from G-code duration and an assumed power draw, and compare with published measured ranges | An energy estimate with sensitivity bars; flag modelled vs measured |
Pick one, write its CLAIMS.md before you start, and set the evidence label you hope to reach. For most of these the honest ceiling is "demonstrated at lab scale".
Why it is hard
Starting is hard for three reasons: the literature is large and ranked by proxies; most evidence is single-lab, so you must judge it yourself; and a good problem is not the one with the most papers. The last point is where counts mislead most: a crowded topic is not an open problem.
Open questions
- Open APIs change. OpenAlex's response now warns that unauthenticated use shares a small daily budget with others on the same network, and Semantic Scholar answered with HTTP 429 (too many requests) when we repeated a search; plan for keys and for caching.
- There is no shared way to mark how well a claim is evidenced. The labels in this book are ours, not a standard.
Industry impact
Engineers who can read and reproduce research can check a claim before relying on it, which shortens the path from a published result to a decision. Whether a regulator accepts such evidence is the subject of Chapter 12.
Choosing a problem
Do not choose from this chapter alone. The next and last chapter, The honest verdict, ranks the unsolved problems, with the evidence for each ranking (and says that the ranking is the author's synthesis). Read it with your CLAIMS.md open: a good first project is small enough to finish in a few weeks, touches one of the higher-ranked problems, uses data you can obtain, and states the evidence label it could reach.
When it breaks
Real failures from preparing this chapter:
- Empty
cite.mjs searchwith a high--min: lower--min, drop the date filter, or reword. Fuzzy relevance means the paper may not be in the top 150. - HTTP 429 from Semantic Scholar on repeated searches and on
cite.mjs abstract(OpenAlex is the fallback). Wait, cache results, or use a key. nullauthor id in OpenAlex crashed our firstselfcitescript (Cannot read properties of null); the script now skips authors without an id and reports how many it used.- Irrelevant top hit from OpenAlex free-text search. Read titles; use a DOI where you can.
make: not foundin a slim Docker image. Callpythondirectly inCMD.- Fresh installs of numeric libraries are slow (about 2 minutes here, both for a venv and for a Docker build); cache wheels or reuse the environment.
What to try next
- Run
litcheck.mjs countson the references of a paper you trust and see which counts disagree most. - Build a reading list of 20 papers: 5 foundational, 10 recent, 5 critical or negative results.
- Fill a
CLAIMS.mdfor a paper you disagree with, and list what evidence would change your mind. - Copy the template, put Lab 5's analysis in it, and have a colleague reproduce it from a clean machine.
- Choose one of the ten projects and write its data card before touching the data.
Why it matters
Reading well and starting reproducibly is the difference between adding to the literature and adding to the noise. It is also the habit that the rest of this book rewards: asking how strong the evidence is.
In the real world
- AM Bench publishes measurements, challenges and outcomes so that groups can be compared on the same data [10][12].
- ASTM F42 and ISO/TC 261 coordinate standards for terminology, test methods and data [13][14].
- Dataset records with DOIs, such as the NIST record used in Lab 5, let others cite and re-use data [11].
Key terms
- Foundational paper — a widely cited, usually older paper that defines or reviews a topic.
- Forward and backward snowballing — following a paper's references (backward) or the works that cite it (forward) to find related literature.
- Age bias — the tendency of cumulative citation counts to favour older papers.
- Self-citation — a citation where the citing and cited works share an author.
- Evidence level — this book's label for how strongly a claim is established: lab scale, production or contested.
- Unit of replication — the independent thing that was repeated (a build, a specimen, a lab), as opposed to repeated measurements of one thing.
- DOI — digital object identifier, a persistent link to a paper or dataset.
- Data card — a short document describing a dataset's source, licence, structure and limitations.
- FAIR — findable, accessible, interoperable, reusable: a set of principles for data.
- Pinned environment — software versions recorded exactly, so a run can be repeated.
Where this leads
Chapter 19, The honest verdict, weighs the whole book, ranks the problems industry struggles with most, and gives advice on choosing one. Bring your CLAIMS.md and your shortlist of project ideas.