The Problem: An Obsolescent Enigma
The factory roared in perpetual agony, its belly devouring metals, melting them into primordial chaos. At the heart of this relentless transformation stood the refractory linings of the smelting vessels, their duty noble yet thankless: to withstand the crucible’s fury. They eroded, flaked away into dust, sacrificed to the molten abyss.
Engineers pondered: Could the wear of these refractory linings be foretold, preempted, controlled? For years, they relied on intuition, manual inspections, and post-mortem assessments of used vessels, but uncertainty ruled. The collapse of a lining could spell disaster—a shutdown, a spill, a catastrophe. A method was needed, an oracle to divine the future of the furnace walls.
The Solution: The Machine’s Verdict
Into this arena stepped Refractory Intellectual Property GmbH & Co. KG, bearing an invention of algorithmic prophecy. Their patent application EP 2789960 promised an automated method for determining the condition of a metallurgical smelting vessel’s refractory lining. It proposed a system of data gathering—tracking the vessel’s materials, wall thickness, exposure to heat, slag composition, operational temperatures, and usage cycles. These variables would be harvested, compiled into a computational model, and subjected to a regression analysis.
From this, the machine would learn. It would adapt, refine its understanding, and calculate the wear of the refractory lining. No longer would the walls crumble unpredictably—now, they would reveal their decay in the cold certainty of numbers.
The First Refusal: A Machine That Cannot Be Built
The European Patent Office (EPO), the bureaucratic sentinel of innovation, stood unimpressed. On July 29, 2021, the Opposition Division rendered its judgment: the invention was inadmissible. Article 83 EPC demanded that an invention be disclosed in such a way that a person skilled in the art could execute it. This, they declared, the applicant had failed to do.
“The patent provides no concrete model, no executable formula,” they reasoned. “It describes data collection and machine learning in the abstract, but where is the model? Where is the training data? How is the regression analysis configured? The skilled person is left adrift, burdened with the task of constructing the very thing the patent claims to have invented.”
The patent, thus, was revoked. The ruling had been given. The invention had been sentenced to oblivion.
The Appeal: A Desperate Plea for Redemption
Refusing defeat, Refractory Intellectual Property GmbH & Co. KG lodged an appeal. They implored the Technical Board of Appeal 3.2.03 to reconsider. Their arguments resounded in the tribunal’s chamber:
- The invention was not an unworkable abstraction but a manifestation of modern machine learning principles.
- The model was inherently self-learning; it did not require exhaustive specification.
- The skilled person would know how to implement such an approach based on available AI methodologies.
A verdict was sought. Would the Tribunal restore the patent to life, or was its fate sealed?
The Board’s Final Judgment: The Sentence Confirmed
The Board of Appeal, on July 23, 2024, delivered its final ruling. The patent, it declared, could not stand. The invention failed to meet the standard of sufficient disclosure under Article 83 EPC.
Their reasoning was labyrinthine, their arguments inexorable:
- A Model That Did Not Exist: The patent spoke of a “computational model” but failed to define its architecture, let alone provide a working example.
- Regression Without Direction: The concept of “regression analysis” remained a ghostly whisper, with no methodology detailed. Was it linear? Logistic? Bayesian? The patent did not say.
- No Starting Point, No Proof: Unlike prior cases where a machine learning model had been at least outlined, here, no structure was provided. The skilled person was abandoned in the void.
- A Research Program in Disguise: Instead of an invention, the patent demanded the reader embark on an odyssey of discovery, to figure out for themselves how to make the system work.
The Appeal Board dismissed the appeal. The request for reimbursement of the appeal fee was denied. The invention was erased from the register of granted patents, cast into the abyss of unpatentable ideas.
The Lesson: Future Navigations Through Bureaucratic Darkness
The judgment, though damning, was instructive. Future applicants were left with a warning and a map to avoid similar doom:
- Do not describe a machine that learns—describe how it learns. A patent must include explicit guidance on its training data, model architecture, and parameters.
- Regressions must not be a mystery. Define the statistical methods, their application, and their expected outputs.
- Provide examples. A working prototype or demonstrable embodiment is the surest defense against the void of insufficiency.
- Avoid the pitfall of vagueness. If a “skilled person” must undertake extensive experimentation to understand the invention, it is no invention at all.
Thus ended the saga of EP 2789960, a vision of AI-driven metallurgy swallowed by the yawning chasm of legal formalities. The furnace walls would continue to wear, to decay in quiet oblivion, until another would dare to file again—this time, armed with the wisdom of the tribunal’s decree.
The judgment had been made. The doors of the Tribunal closed. The smelting continued.
Based on T 1669/21 23-07-2024.