Energy per Good Casting: The KPI That Prevents False Savings in Foundry AI
Short answer: In a stable casting process, lower energy per shot is not proof of a saving. The defensible KPI is energy per accepted casting, evaluated inside the approved process window and checked against scrap, rework, part mass, and production mix. That is how a small AI fine-tuning step becomes an economic gain rather than a misleading efficiency signal.
Energy per Good Casting: The KPI That Prevents False Savings in Foundry AI
Foundry teams are usually very good at reaching a process that works. They tune metal temperature, die or mould conditions, filling, pressure, cooling, lubrication, and cycle sequence until the casting meets its requirements. Once the process is stable, the rational instinct is to stop touching it.
That caution is valuable. But “stable” does not automatically mean “economically optimal.” There may still be a small, safe adjustment that reduces cycle time or energy consumption while keeping quality inside the approved envelope. Artificial intelligence can help search this narrow space. The key is choosing a target that cannot reward a false saving.
Why kWh per shot can lie
Energy intensity is energy divided by a defined production output. ISO 50006 treats energy performance indicators and baselines as tools that must fit the organisation’s actual energy use. The U.S. Department of Energy similarly defines energy intensity as energy use per unit of activity or output.
For a casting line, “one shot” is easy to count—but it is not always the output that earns revenue. If a trial lowers energy per shot while increasing scrap or rework, the plant may consume more energy for every accepted casting delivered.
A practical primary KPI is:
Energy per good casting = energy inside the agreed boundary / accepted castings produced
For multi-cavity tools, varying part mass, or changing product mix, use a denominator that represents the real output: accepted cavities, accepted kilograms, or an agreed product-mix-normalised unit. The definition must remain unchanged between baseline and trial.
Define the measurement boundary before optimisation
The European Commission’s foundry BAT conclusions define foundry-specific energy consumption against liquid metal output and include energy used for melting, holding, transfer and ladle preheating within the stated boundary. Your operational KPI may be narrower or wider, but it must be explicit.
For example, decide whether the comparison includes:
melting and holding;
furnace-to-machine transfer and ladle preheating;
casting machine, hydraulic or pneumatic demand;
die or mould thermal control;
cooling, extraction, compressed air, lubrication, trimming, or heat treatment when the tested change affects them.
Then synchronise meter data with the relevant shot, batch, alloy, tool, cavity, shift, and final quality disposition. A beautifully precise meter attached to the wrong production record is still poor evidence.
A seven-step gate for safe fine-tuning
Freeze the quality requirements. Keep dimensional, metallurgical, leak, mechanical, and visual acceptance criteria unchanged.
Choose one primary outcome. Use energy per accepted casting or cycle time per accepted casting—not a loose collection of favourable signals.
Build a representative baseline. Cover normal warm-up, steady production, alloy or tool states, ambient conditions, and realistic micro-stoppages.
Normalise relevant variables. Account for part mass, product mix, scrap disposition, idle periods, and any factor that materially changes energy use.
Let AI search only inside engineering limits. Do not allow a recommendation to cross validated temperature, pressure, fill, solidification, or equipment constraints.
Change one controllable setting at a time. Hold the remaining conditions as steady as production allows and log the recommendation, operator decision, and result.
Accept only a repeatable net gain. The improvement must survive a comparison of energy, good output, quality, cycle time, and process stability.
This is measurement and verification, not a single before-and-after screenshot. The U.S. DOE’s current M&V guidance emphasises a baseline and adjustments for relevant variables such as production.
Modelled example: lower energy, worse economics
Consider a simplified, modelled scenario—not a Castella customer result:
Baseline: 1,000 cycles, 950 accepted castings, 1,900 kWh.
Baseline KPI: 1,900 / 950 = 2.00 kWh per good casting.
Trial: 1,000 cycles, 900 accepted castings, 1,824 kWh.
Trial energy per cycle is 4% lower.
Trial KPI: 1,824 / 900 = 2.03 kWh per good casting.
The energy meter looks better, but energy per sellable casting is about 1.4% worse. The apparent saving disappeared because the denominator changed.
Where Castella’s two optimisation modes fit
Castella’s Cycle Time Reduction and Energy Cost Reduction modes are most useful after the process is already acceptable. They can rank small candidate changes inside the approved operating window, while the engineer remains responsible for constraints and validation.
In a satisfied, working process, Castella may identify a predicted 5–15% opportunity from a single controlled change. That range is an optimisation hypothesis to test, not a performance promise. The result is accepted only when plant data confirms that:
quality and process capability have not deteriorated;
energy per accepted casting or cycle time per accepted casting improves;
the change is repeatable across a representative production window;
no downstream cost has been displaced into rework, maintenance, tooling, or another energy consumer.
This is the difference between AI as an impressive recommendation engine and AI as a trustworthy decision-support tool.
Evidence limits
No universal parameter change can guarantee the same saving across HPDC, LPDC, gravity die casting, or sand casting. Alloy, geometry, wall thickness, gating, thermal balance, machine condition, furnace efficiency, inspection system, and data traceability all influence the result. A model should expose uncertainty, stay inside engineering constraints, and abstain when the operating state is outside its evidence.
For the preceding step—how to fine-tune a stable process—read Beyond a Stable Casting Process: AI Fine-Tuning for Cycle Time and Energy Cost. To discuss how the measurement boundary and validation gate should fit your line, visit Castella AI for Die Casting.
Frequently asked questions
Is energy per shot useless?
No. It is a useful diagnostic and control signal. It becomes misleading only when it is treated as the business result without checking accepted output and quality.
Should the denominator be good parts or good kilograms?
Use the unit that represents your production economics. Good parts work for a stable single product. Accepted kilograms or a product-mix-normalised unit may be better when mass and product mix vary.
Can cycle time and energy be optimised together?
Yes, but define a primary objective and hard quality constraints. A shorter cycle that raises scrap, rework, die temperature instability, or downstream energy use is not a verified gain.
Does a 5–15% predicted opportunity mean guaranteed savings?
No. It is a candidate improvement range generated within the model and process constraints. It must be validated on the actual line with one controlled change and an agreed measurement-and-verification plan.
Primary references
Serdar Perçin writes about metallurgy, casting-process engineering, and verifiable industrial AI. The engineering decision remains with the foundry team; the model should make that decision better informed, not less accountable.
Comments