The honest verdict: what industry still cannot do
📍 Where we are: Part VII · The Verdict — Chapter 19 of 19. The research starter kit showed how to read and reproduce research. This last chapter weighs all eighteen chapters, including the two labs, and says plainly what industry still cannot do.
After a long expedition, a good guide draws the map in three colours. Green: roads that carry traffic every day. Yellow: tracks someone has walked once. Red: places where even the experts disagree about what is there. Then she says which red patches block the most roads. This chapter is that map, drawn from the book's own evidence. The ranking of the red patches is her judgement, not a vote.
What this chapter covers
No new research. We sort what the earlier chapters established into a ledger, add what the two labs showed, name the recurring tensions, rank the problems industry struggles with most, and advise on choosing one. Where a chapter dropped or softened a claim, this chapter does too. The ranking is ours and is flagged as such.
Problem
A verdict can fail by repeating vendor optimism or by dismissing real progress because the physics is hard. The preface offered the remedy: ask how strong the evidence is and how close it is to a factory, and keep three senses of "solved" apart: we understand the mechanism, a recipe works, or an authority accepts the evidence. Most AM problems are solved in the first or second sense and open in the third. That gap is why blind benchmarks exist; the 2022 NIST AM Bench round drew 138 sets of modelling submissions, up from 46 in 2018 [1].
State of the art
The ledger
"Production" is routine commercial use; "pilot" is lab scale, a testbed or a single case; "not shown" is a claim the chapters could not support; "contested" means credible sources disagree.
| Area (chapter) | Production | Pilot | Not shown or contested |
|---|---|---|---|
| Melt pool (1) | Vendor parameter sets from density cubes | Synchrotron-labelled pore detection | New-alloy window without trial builds; thin-plate thresholds on production machines |
| Defects, stress (2) | Preheat, stress relief, HIP | Nucleants against cracking | Quantitative crack rule; distortion accuracy on complex parts |
| Microstructure (3) | Frozen recipe, HIP, machining | Site-specific microstructure | Properties by location; fatigue parity with wrought |
| Simulation, ML (4, 5) | Distortion pre-compensation (maturity unquantified) | ICME chains; Bayesian optimisation | ML closed-loop correction in production; no widely adopted ML benchmark (our assessment) |
| Sensing, control (6, 7) | Monitoring as logging; vendor-reported overhang control [2] | Operando-labelled detectors; feedback testbeds | Monitoring replacing CT; machine-to-machine transfer |
| Alloys (8) | Four mature alloys | GRX-810; refractory alloys | Crack-free refractory alloys at scale; universal printability rule |
| Multi-material (9) | Print one, join the rest | Voxel-level soft matter [3]; graded metals | Industrial electronics and 4D parts; sharp versus diffuse interfaces |
| Bioprinting (10) | Acellular implants | Perfusable tissue; volumetric printing | Bioprinted organs; regulatory route |
| Large scale (11) | Pellet-printed moulds | WAAM bridge, a one-off [4] | Routine reinforced concrete structures |
| Qualification (12) | NASA, FAA, EASA, FDA, ISO/ASTM routes [5] | Monitoring-assisted qualification | Monitoring alone for flight-critical parts (none found); machine versus part qualification |
| Digital thread (13) | Maturing formats | Data-pedigree proof of concept | Seamless cross-company thread; size of spare-parts niche |
| Sustainability (14) | Measured printing energy (wide spread) | Powder-reuse LCAs, few builds | "AM is greener" as blanket claim; modelled use-phase credits |
| Cost, scale (15) | LEAP nozzle tip; aligner molds | Rotary multi-laser concept | Linear gain from more lasers; post-processing share |
Nearly every Production entry is a workaround (freeze the recipe, add HIP, flag instead of decide, consolidate parts), not a solved physical problem. The Pilot column is long, and most entries are one machine, one alloy, one laboratory.
Ten problems ranked by the author's synthesis, not by a published consensus: chapters touched, status of the best proposed solution, and the workaround in use today.
Original diagram by the authors, created with AI assistance.
What the two labs added
Chapters 16 and 17 ran checks many papers skip. They are small single-dataset runs, so they illustrate a failure mode, not its frequency.
- Calibration is not validation (Lab 4). NIST measured the AMB2018-02 case B melt pool at 123.5 µm wide and 36 µm deep [6]. AdditiveFOAM's liquidus isotherm gave 123.8 and 36.2 µm, but the tutorial's absorption and source dimensions were calibrated on that same benchmark data [7], so the match is not independent validation. The solidus isotherm, an equally defensible fusion line, was 14% too wide and 21% too deep, and plausible absorptivity and conductivity values moved Rosenthal's width from 136 to 180 µm; mesh error was under 3%.
- The split decides the answer (Lab 5). On NIST single tracks [8], 7 process settings measured 6 times each, a Gaussian process had an RMSE of 2.2 µm under a random split and 40.8 µm when a whole setting was held out (the no-skill mean: 48.3 µm). Its stated 95% intervals covered 79% of held-out rows. Leakage is a documented cause of overstated ML results across fields [9]. The data are bare plates from one testbed, so nothing carries over to powder beds or other machines.
Neither lab says modelling is unreliable; they show that a good match raises two questions: what was tuned on the same data, and what the split respected.
Why it is hard
- Speed against quality. Pushing power or speed brings keyholing, lack of fusion and spatter; more lasers raise output less than linearly [10].
- Proxy against defect. Sensors see surface light and heat; one detector found flaws of 200-1000 µm, while fatigue-relevant flaws are tens of micrometres [11].
- Calibration against prediction. A tuned model is not a predictor (Lab 4); in AM Bench 2022 the most sophisticated code was not automatically the most accurate [1].
- Validation against convenience. Random splits flatter; held-out settings, machines and labs are honest (Lab 5). Ground truth is expensive, so demonstrations outnumber validations.
- Freeze against improve. Qualification takes extensive testing, millions of dollars and years, and minor changes mean re-qualification [12], so processes are frozen.
- Data that does not travel. Records of powder, process and inspection are fragmented across vendors [13][14].
Open questions
The ranked list
This ranking is the author's synthesis, not a published consensus: no industry survey or source in this book ranks these problems. Criteria, judged from the chapters: leverage (how many other problems depend on it), persistence (how long it has resisted effort, and whether authorities lack an agreed rule) and consequence (cost or blocked adoption reported). The book leans on laser powder bed fusion, so another author could reorder it. Ranks 1 to 3 should be more stable than 4 to 10, which are close calls.
| Rank | Problem | Evidence and why this rank |
|---|---|---|
| 1 | Qualifying parts without exhaustive re-testing (Ch 12) | Thousands of tests, millions of dollars, years [12]; machine-level versus part-level routes are contested; no regulator found accepting monitoring alone. Highest leverage: other chapters end in "who accepts this evidence?". |
| 2 | Trustworthy in-situ detection and control (Ch 6, 7) | Detection shown at 200-1000 µm [11]; labelled detection on tracks and thin walls [15]; feedback on testbeds [16]; in-process evaluation at lower readiness than post-process methods [17]. The route to rank 1; validation is weak. |
| 3 | Predicting fatigue from structure (Ch 2, 3) | The worst local defect governs; process and post-processing reduce but do not remove defects [18]; HIP plus finishing is the workaround, parity with wrought contested. Decides safety-critical load. |
| 4 | Validated transfer across machines, powder lots and alloys (Ch 4, 5, 8, 16, 17) | Models are calibrated per machine; powder state is path-dependent; the labs show calibration passing as prediction. Partly a practice problem, hence below 1 to 3. |
| 5 | Post-processing cost and low build rate (Ch 15) | Post-processing strongly influences modelled cost [19], share unquantified; multi-laser gains sublinear [10]; series production where consolidation pays [20]. Incremental. |
| 6 | AM-native alloys in qualified production (Ch 8) | Nucleant aluminium [21] and GRX-810 [22] are lab-scale; four alloys are mature. Long, not blocked; returns to rank 1. |
| 7 | Interoperable digital thread (Ch 13) | About 3000 TB estimated for one qualification [13]; IP and data-quality cost are barriers [14]; no adopted schema. Largely organisational. |
| 8 | Interface integrity in multi-material parts (Ch 9) | Graded paths can cross brittle phases; interface allowables barely exist. Severe but narrow. |
| 9 | Reinforcement and codes for large structures (Ch 11) | Reinforcement still unresolved [23]; qualification principles exist [24] and the ICC reports a US wall standard approved July 2026 [25]. One sector. |
| 10 | Comparable environmental accounting (Ch 14) | Specific energy 1-2 orders above conventional [26]; utilisation changed per-part impact 14-18 times in one study [27]; use-phase savings modelled, equal performance assumed [28]. Shapes procurement, not feasibility. |
Bioprinting and micro-scale AM (Ch 10) are outside the ranking on purpose: cell-laden implants are investigational, whole-organ printing is "not close" [29], and it is not yet an industry struggling at volume.
Industry impact
Advice for choosing a problem
The author's opinion, built from the pattern above.
- Start from the ledger, not the headline. A problem with many pilots and little validation (ranks 2 and 4) is often worth more than another demonstration.
- Ask who would accept your result. For a regulator, report probability of detection and confidence bounds, not only accuracy [11].
- Use ground truth and open data, and name your split. Validate on a held-out setting, build, machine or lab (Lab 5); state what was tuned on the same data (Lab 4).
- Report the noise floor and interval coverage, e.g. a "95%" interval that covers 79%.
- Match scale to leverage. Ranks 1 to 3 need collaboration; a small team can contribute to ranks 4 to 8 with narrow scope, and its honest ceiling is usually "demonstrated at lab scale".
- Engineers choosing a process: ask vendors for independent test data, model unit cost for your own geometry and volume, and look first where consolidation pays [20].
How to get in
Do not choose from this chapter alone. The research starter kit gives the method: write a CLAIMS.md with an evidence label first, use the project template, and pick from ten first projects tied to chapters and labs. Lab 4 and Lab 5 are the quickest way to feel the gap between a result and a validated result. A good first project finishes in weeks, touches a higher-ranked problem, uses data you can get, and states the evidence label it could reach.
Why it matters
An honest map saves effort: it tells a manager where a vendor claim needs independent data and a student where a project would matter. It also shows that AM's successes are real; they come from applications where the weaknesses do not bite.
In the real world
The LEAP fuel nozzle tip succeeded through design consolidation [20], and aligner production prints molds in series [30]; neither solved melt-pool physics. NASA and FAA frameworks freeze a process and control it statistically [31][32].
Key terms
- Ledger — a table sorting claims by evidence level.
- Workaround — a practice that avoids a problem rather than solving it, such as freezing a recipe.
- Leverage — how many other problems depend on solving one.
- Blind benchmark — a test where measurements are held back until predictions are submitted.
- Calibration versus validation — tuning inputs to data, versus checking against data not used for tuning.
Where this leads
This is the last chapter of Book 4 and of the series (start again at Book 1, Book 2 or Book 3, or reread the Book 4 preface). The most useful next step is to pick one rank, read its chapter with the references open, and start a project with the research starter kit.