Edge ML Emulator for Wildfire Spread Prediction
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current environmental computational fluid dynamics (CFD) models are compute-intensive and cannot be deployed on edge devices due to memory, power, and data transmission constraints, limiting their use for in-the-field decision-making, especially during harsh conditions with limited cloud connectivity, such as wildfire response.
Innovation Solution
A machine learning-based framework that generates gridded predictions of wildfire spread by sensing climate and earth surface data, processing historical wildfire data, and training probabilistic mapping function emulators to predict wildfire occurrence and spread, allowing for scalable, cost-effective, and energy-efficient simulations on edge devices.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If process-driven CFD simulation is used for wildfire spread prediction, then prediction accuracy is improved, but computational cost and energy consumption increase excessively
Solution Approach 1:
The patent creates a machine learning emulator that copies the functionality of the complex CFD simulation process. The emulator is trained on CFD simulation data and then used to predict wildfire spread, providing accurate predictions without the high computational cost of running full CFD simulations in real-time
Solution Approach 2:
The patent replaces the mechanical CFD simulation process with a machine learning-based computational approach. The ML model substitutes the traditional physics-based numerical solving process, achieving similar predictive accuracy with significantly reduced computational requirements
2Reliability
If realistic process-based models are deployed, then prediction reliability is improved, but device memory and power constraints prevent deployment on edge devices
Solution Approach 1:
The patent creates a simplified copy of the complex process-based model that runs on edge devices. The ML emulator captures the essential predictive capabilities of the full model while being lightweight enough for deployment on resource-constrained devices with limited memory and processing power
3Measurement precision
If data assimilation is performed in process-based models, then prediction accuracy is improved, but computational time and resources increase significantly
Solution Approach 1:
The patent performs data assimilation and model training in advance during an offline phase. The ML emulator is pre-trained on historical data and CFD simulations, so that during real-time prediction, no additional data assimilation computation is needed, enabling fast predictions with pre-integrated knowledge
4Ease of operation
If cloud connectivity is limited during harsh conditions, then in-the-field decision-making capability deteriorates, but process-based models require cloud connectivity for computation
Solution Approach 1:
The patent enables edge devices to perform wildfire predictions autonomously without requiring cloud connectivity. The ML emulator runs locally on the edge device, allowing first responders to make informed decisions in the field even when disconnected from cloud infrastructure during harsh conditions
Data Source
AI summary
Apparatuses, methods, and systems for generating gridded predictions of a probability of wildfire spread for geographies are disclosed. One method includes sensing observational climate and earth surface data, obtaining historical data on wildfire spread events, obtaining gridded climate data, creating a set of input features, creating a gridded wildfire data set, training a model that learns one or more probabilistic mapping function emulators between the set of input features and the gridded wildfire data set, which predicts a first probability of wildfire occurrence and a rate and extent of wildfire spread within a geographical region and at a specified period of time, and generating gridded wildfire prediction data including a second probability of wildfire occurrence and spread within the geographical region, using the model and a new set of input features over the geographical region but for a different time period.