A Leading Global Industrial Manufacturer Accelerates HVAC Filter Design with AI-Driven Surrogate Modeling

The Design Iteration Bottleneck in Industrial HVAC Engineering

A leading global industrial manufacturer needed to improve the design of a high-performance HVAC filter to gain competitive advantage. The engineering challenge is fundamental: balancing airflow efficiency and pressure drop (ΔP) - a critical KPI that directly impacts energy consumption. Changes to porous media geometry that improve one metric typically degrade the other, so optimal design requires extensive exploration of the design space.

‍

With conventional CFD simulation, each new filter design required full computational fluid dynamics analysis taking days per iteration. Engineers could evaluate only 5 to 10 designs per week, restricting both innovation and optimisation. Reliance on high-performance computing resources made routine validation expensive and difficult to scale. Sequential design-simulate-analyse-iterate workflows meant engineering teams were spending most of their time validating designs rather than driving improvement.

‍

A Hybrid Deep Learning Architecture That Evaluates Designs in 30 Seconds

The Key Ward team developed a hybrid AI-driven surrogate modeling workflow built on three components: complex CAD geometries are converted into structured 3D data using signed distance functions (SDFs), creating consistent inputs for machine learning models; a 3D convolutional neural network (CNN) shape encoder learns geometric features from filter shapes combined with physical parameters including thickness and material type; and a multi-layer perceptron (MLP) predicts pressure drop (ΔP). The surrogate model is embedded directly into the design environment for real-time feedback.

‍

Technical validation results: 91.9% R² score on validation data with a root mean square error (RMSE) of 16.6 - achieved with only a modest training dataset. Business impact: evaluation time reduced from hours and days to approximately 30 seconds per design, and the team shifted from 5–10 designs per week to thousands of evaluations per day. Parallel exploration replaced sequential testing.

‍

What the Full Case Study Documents

The full case study documents the complete surrogate model architecture - the SDF geometry encoding, the CNN shape encoder and MLP prediction structure, training methodology, and full validation results with accuracy metrics. It includes the measured speed improvement, expansion in daily evaluation capacity, and cost reduction in HPC reliance. It also covers the scalable framework extendable to new geometries, materials, and additional KPIs. If your industrial engineering team is constrained by CFD turnaround time during design exploration, the approach documented here is directly applicable.