Case: T 0048/24 | Board: Technical Board of Appeal 3.5.06 | Date: 13 November 2025
Key Takeaways
1. Claiming broad ML outputs without technical detail is fatal. A patent that merely states a result to be achieved — without disclosing how a machine learning model should be implemented, trained, or evaluated — will not survive a sufficiency challenge, especially when the claimed output spans a wide range of possibilities.
2. “One way” of working the invention is not enough if the claim is broad. Even accepting that a coarse three-label waste classification could be implemented using a CNN, this did not enable the skilled person to carry out the full breadth of the claim, which extended to many other output types and precision levels.
Summary
Hitachi Zosen Inova AG successfully challenged a patent held by Ebara Environmental Plant Co., Ltd. covering a machine learning device for estimating the composition of waste stored in a pit — information relevant to controlling incineration plants. The Opposition Division had initially rejected the challenge, but the Board of Appeal reversed that decision and revoked the patent.
The core issue was sufficiency of disclosure under Article 100(b) EPC. The patent’s claim was broadly drafted: it covered any “data of” a captured image as input and any “value representing composition” as output, without limiting the type of ML model, the training approach, or the required accuracy. The Board found this combination of extreme breadth and minimal technical disclosure fatal. The patent offered no concrete working example, no guidance on which input data or model architectures were suitable, and no information on achievable accuracy — leaving a skilled person to conduct what would amount to an open-ended research programme across an enormous parameter space. That constitutes undue burden. All eight auxiliary requests failed for the same essential reason, and the patent was revoked in full.