ML Model Training with Hardware Error Simulation for Datacenter Reliability

Resolve Bottlenecks,
Find Innovative Solutions
Generate Solutions

Solution Overview

Problem

Existing machine learning models struggle to maintain accuracy and efficiency when deployed in datacenters with hardware component failures, as current error prediction and redundancy methods either fail to accurately predict errors or require additional resources and space.

Innovation Solution

Train machine learning models to handle errors by simulating component failures during training, using historical error data to determine error frequencies and durations, and updating model parameters to compensate for errors, thus enabling the model to perform accurately even with hardware failures.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Reliability

If built-in redundancies are implemented to mitigate hardware failures, then system reliability is improved, but device complexity and resource consumption increase

Engineering Contradiction:
Improvesystem reliabilityVSAvoiddevice complexity
Core Design Contradiction:
ReliabilityVSDevice complexity

Solution Approach 1:

The patent applies this principle by training the machine learning model to actively learn from and adapt to hardware errors during training. Instead of preventing errors with redundant components, the model is exposed to simulated hardware failures and learns to compensate for them, converting the harmful effect of errors into a beneficial training signal that improves robustness without adding physical redundancy

Inventive Principle:
Principle #22Blessing in disguise (Convert harm into benefit)

Solution Approach 2:

The patent implements self-service by enabling the machine learning model to autonomously detect and correct its own performance degradation caused by hardware errors. The model uses its internal representations and learning capabilities to adapt to errors without external intervention or additional hardware components, making the system self-healing rather than relying on external redundancy

Inventive Principle:
Principle #25Self-service

2Reliability

If checkpoint creation is used to handle hardware failures, then reliability is improved, but loss of time and productivity increase due to training progress loss

Engineering Contradiction:
ImprovereliabilityVSAvoidtraining progress loss
Core Design Contradiction:
ReliabilityVSLoss of time

Solution Approach 1:

The patent applies preliminary action by pre-training the model with simulated hardware errors included in the training process. This prepares the model in advance to handle actual hardware failures during deployment, eliminating the need for checkpoint-based recovery and allowing continuous operation without training progress loss when errors occur in production

Inventive Principle:
Principle #10Preliminary action

3Reliability

If error prediction processes are implemented, then reliability is improved, but device complexity and energy consumption increase

Engineering Contradiction:
ImprovereliabilityVSAvoidenergy consumption
Core Design Contradiction:
ReliabilityVSUse of energy by moving object

Solution Approach 1:

The patent converts the harmful effect of hardware errors into a beneficial training signal by incorporating error simulation directly into the training process. This eliminates the need for separate error prediction systems, reducing energy consumption while maintaining reliability through the model's learned error compensation capabilities

Inventive Principle:
Principle #22Blessing in disguise (Convert harm into benefit)

4Reliability

If redundant hardware components are deployed, then reliability is improved, but loss of substance and resource consumption increase

Engineering Contradiction:
ImprovereliabilityVSAvoidresource consumption
Core Design Contradiction:
ReliabilityVSLoss of substance

Solution Approach 1:

The patent uses virtual copying of error conditions through simulation during training, rather than physical copying through redundant hardware. The model learns from simulated error scenarios that replicate real hardware failure modes, achieving the same reliability improvement without the resource cost of physical redundancy

Inventive Principle:
Principle #26Copying

Solution Approach 2:

The patent changes the operational parameters of the training process by introducing error simulation parameters that control the type, frequency, and severity of simulated hardware failures. This allows the model to learn robustness to various error conditions without changing the physical hardware configuration or consuming additional resources

Inventive Principle:
Principle #35Parameter changes

Data Source

PatentEP4697235A1Machine learning model training with hardware error simulation
Publication Date: 2026.02.18 GOOGLE LLC
  • EP4697235A1 patent drawingFigure 1
  • EP4697235A1 patent drawingFigure 2
  • EP4697235A1 patent drawingFigure 3

AI summary

Methods, systems, and apparatus, including computer-readable storage media for handling component error of datacenter infrastructure by training machine learning models to handle errors when the models are deployed. Rather than predict or mitigate errors in the hardware of a datacenter or another site deploying the machine learning model, the model is trained to perform a task with comparable accuracy and efficiency even when some hardware on which the model is deployed fails. Component error can instead be simulated during training to cause the machine learning model deployed on the infrastructure to learn to correct errors caused by the component errors. The model can continue to be trained to compensate for periods in which not all parts of the model are available at inference. Updates to the model can be backpropagated to correct errors for handling instances of component error.