When One Parameter at a Time Is Not Enough
Short answer: Changing one parameter at a time is useful for confirming a low-risk adjustment, but it cannot reveal interactions between casting variables. When the effect of one setting depends on another, foundries need a bounded design-of-experiments (DOE) gate before an AI recommendation becomes a production recipe.
A casting process can look stable and still contain hidden combinations. Vacuum may help only above a certain intensification pressure. A higher second-phase velocity may improve filling while also increasing air entrainment unless venting capacity is adequate. Sand moisture can behave differently at different mold hardness or permeability. If we move only one knob and stop after the requested quality is reached, we may miss the more economical operating point—or attribute the improvement to the wrong cause.
This is where AI can add value, but only inside an engineering-controlled experiment.
Why one-factor-at-a-time trials have a limit
One-factor-at-a-time, often called OFAT, is understandable on the shop floor. It is easy to explain, limits disturbance, and can be appropriate for a final confirmation: change one bounded setting, observe the next production block, and check that quality guardrails remain intact.
The limitation is structural. OFAT estimates a local effect while other factors stay at their current values. It does not tell us whether the same effect survives at a different die temperature, vacuum level, melt condition, sand batch, or thermal state. In statistical language, it does not efficiently estimate interaction terms.
The NIST definition of experimental design includes deliberate changes to one or more process factors and models that can contain interactions. NIST also notes that design choice depends on the objective and factor count; screening designs answer different questions from response-surface designs used for interactions, curvature, optimization, and robustness.
The practical distinction is simple:
Use a single change to confirm a narrowly defined, low-risk local recommendation.
Use DOE when several plausible causes may combine, when the response is curved, or when the operating window must be optimized rather than merely repaired.
What a parameter interaction looks like in casting
Consider a modeled HPDC example. Historical data suggests that stronger vacuum is associated with lower porosity. A simple model ranks vacuum as the most important controllable factor.
At low intensification pressure, however, the vacuum change produces little repeatable benefit. At a higher but approved pressure, the same vacuum change produces a clear improvement. Vacuum was not “wrong”; its effect depended on pressure.
The same logic appears across processes:
In HPDC, plunger velocity, vacuum, intensification pressure, gate condition, and die thermal state can interact.
In LPDC, pressure-curve timing, metal temperature, die temperature, cooling sequence, and feeding geometry can interact.
In sand casting, moisture, compactability, permeability, mold hardness, pouring temperature, and time to pour can interact.
A safe DOE gate for foundry AI
A useful DOE is not “let the algorithm try everything.” It is a small, bounded learning loop with explicit stop rules.
1. Prove that the process and measurement system are ready
Before testing parameter effects, confirm that the response can be measured consistently and that the process is stable enough to learn from. NIST process-improvement guidance treats a capable measurement system and a stable process as prerequisites. If X-ray classification, leak testing, dimensional measurement, temperature sensing, or scrap coding is inconsistent, the experiment will optimize noise.
2. Define one primary response and hard guardrails
Choose the main objective—porosity, leak rate, fill defect, tensile property, dimensional stability, cycle time, or energy per good casting—and define what must never be traded away. Guardrails may include maximum defect severity, tool-temperature limits, peak pressure, metal-temperature range, safety interlocks, and customer-critical dimensions.
3. Screen factors inside approved limits
Use process knowledge and historical data to shortlist factors. A fractional factorial or carefully selected Taguchi design may screen main effects economically. The exact design depends on the question, factor levels, production constraints, and interactions that must not be confounded.
AI can rank candidates and suggest a compact matrix. The process engineer must approve the factors, levels, forbidden combinations, and stop conditions.
4. Block what production cannot randomize away
Foundry data is time-dependent. Tool, cavity, alloy lot, sand batch, shift, start-up state, and die thermal history can masquerade as parameter effects. Record them and block the experiment where appropriate. Randomize run order only where it is safe and operationally possible.
Without blocking and traceability, an apparent “AI improvement” may simply be a warm die, a new melt, or a different cavity.
5. Add replication, center points, and confirmation
Replicates estimate normal process variation. Center points help reveal curvature. After screening, a focused factorial or response-surface design can investigate the small number of important factors. Finally, return to a confirmation block at the proposed setting and compare it with a suitable baseline under comparable conditions.
The release decision should require repeatability, not one favorable shift.
6. Keep the human approval gate
The NIST AI Risk Management Framework calls for objective, repeatable testing and evaluation, documented limitations, production monitoring, and domain-expert involvement. In a foundry, that means AI proposes and ranks; engineers approve; the machine remains within validated limits; and the evidence behind every recommendation is preserved.
Where Castella fits
Castella can make this loop faster without pretending that correlation is causation. It can help unify shot, cycle, melt, tool, cavity, defect, and energy records; rank plausible controllable factors; surface suspected interactions; and prepare a bounded trial plan with traceable assumptions.
The workflow is deliberately conservative:
diagnose the loss mechanism and data quality first;
propose the smallest informative test;
enforce engineering limits and approval;
compare confirmation production against the baseline;
retain the model version, data window, settings, and quality outcomes.
This extends our article on fine-tuning cycle time and energy cost. Once quality is stable, optimization should also use energy per good casting, not energy per machine cycle. Castella’s HPDC and LPDC optimization approach keeps the recommendation tied to process physics and production evidence.
Evidence limits
DOE improves what can be learned from a limited number of trials, but it does not remove engineering judgment. Results may not transfer between alloys, dies, machines, cavities, seasons, or foundries. A statistically significant term may still be operationally too small, unsafe, or uneconomic. Conversely, an important mechanism may remain hidden if the chosen factor range is too narrow or the response is poorly measured.
Treat the final setting as valid only for the tested operating window. Monitor it after release, especially after tool maintenance, material changes, sensor replacement, or seasonal thermal shifts.
Frequently asked questions
Is one-factor-at-a-time always wrong?
No. It is useful for troubleshooting a single suspected cause and for final confirmation of a low-risk local change. It becomes weak when interactions, curvature, drift, or multiple competing objectives are likely.
Does DOE require stopping production?
Not necessarily. Many experiments can be embedded in planned production blocks with approved limits, blocking, traceability, and stop rules. The design must respect quality, safety, tooling, and delivery constraints.
Can AI choose the experiment automatically?
AI can rank factors and propose an efficient design, but it should not autonomously change production recipes. Engineers must approve factor ranges, forbidden combinations, responses, and release criteria.
How many trials are enough?
There is no universal number. It depends on the objective, number of factors and levels, expected interactions, process variation, replication needs, and acceptable risk. A screening design and a confirmation plan should be sized together.
Comments