Aerospace applications of deep machine learning

Working on problems that don’t yield to physics-based models, or physics-based models take too long to run?

Simulation and modeling have evolved considerably in the past 40-50 years, since computers first began to be used for such purposes. Simulations today run thousands of times faster and with greater mechanistic detail than when we landed on the moon in 1969. Yet there remain certain kinds of problems that are not yet solvable via simulation, as well as some where simulation can work, but the run times are too long to be practical.

Machine learning models have also evolved:

Advanced machine learning (“artificial intelligence”) methods have been applied to highly complex problems, such as:

  • Predicting aerodynamic force and moment coefficients in milliseconds with close-to-CFD accuracy
  • Predicting jet engine compressor mapping to identify when maintenance is required
  • Identifying/classifying missiles from radar tracking data

Want to optimize a design but modifications take too long to evaluate? Or there is no mechanistic solution available? Turn your data into very fast and accurate models:

Computerized design optimization starts with an initial guess, or baseline design, and then some number of design parameters are systematically adjusted to work toward an optimum solution. As each parameter is adjusted, the new design must be evaluated. If that new design requires an inordinate amount of computer time (e.g., CFD) or if there is no mechanistic solution in existence (e.g., solid rocket propellant formulation), then true optimization is prevented, or at best restricted, and less efficient designs result.

Simulations Plus has repurposed its advanced machine learning technology, originally developed and proven best-in-class for molecule property prediction in pharmaceutical chemistry, to enable rapid development of custom applications for a wide variety of aerospace and other problems.

AEROModeler Provides…

A platform for customization to meet a wide variety of modeling needs.

Turn your data into very fast and accurate models

You spend a lot of time and money generating data. Get a second return on your investment by turning your data into fast, accurate, predictive models. Machine-learning models store data efficiently and provide for rapid retrieval and automatic interpolation between observations in many dimensions.

Your application designed for your workflow

Our modeling platform can be tailored to read your data files in your format, provide graphic and tabular outputs of various kinds for your reports and presentations, and turned into an optimization engine to optimize designs using the parameters of your choice to maximize some performance metric while meeting required design constraints. Let us design, develop, and support a turnkey application designed to your specifications, with ongoing review by your team to ensure that the final product will best meet your needs.

Routine Applications

Several prototype applications have been developed using the AEROModeler platform:

Aerodynamic coefficient prediction for missile design optimization

Starting with aerodynamic force and moment coefficients calculated by the AERODSN program from Redstone Arsenal to create a training dataset, we trained advanced artificial neural network ensemble models for axial force, normal force and moment coefficients using a small faction of the dataset and then predicted the remaining values with Q^2 values exceeding 0.99 for all flight regimes (subsonic, transonic, and supersonic). The final CN model predicted 3.1 million transonic data points (different missile configurations and different Mach and angle-of-attack) in 280 sec (>11,000 predictions/sec) on a laptop computer.

Missile diameter recognition from selected trajectory data:

The AERODSN 6-DOF missile trajectory simulation code was used to generate 3,133 unique missile configurations based on 36 design parameters and trajectories with diameters ranging from 0.3-1.5 meters. 15% of the dataset was held out from training as an external test set.

A series of artificial neural network ensemble models was trained to determine the optimum number of independent variables and network neurons to produce the most accurate prediction of missile diameter from the observed trajectory data (eight parameters for the observed time from launch, velocity, altitude, and downrange distance from launch point at motor burnout and at apogee). The final model predicted missile diameters with Q^2 =0.963, and within 5% error at 95% confidence interval and 5.2% error within a 99% confidence interval.

Jet engine compressor mapping:

Data obtained from each of the four engines of a heavy transport aircraft, recorded from multiple aircraft over a period of several days, during which over 94,000 data points were collected. The dataset included aircraft operation at different airfields in the U.S> and abroad, and over a range of operating conditions including climb, cruise, descent, landing, air-to-air refueling, high-speed descents, and multiple airdrop profiles. The dependent variable modeled was the engine pressure ration (EPR). After training on 20% of the dataset, the remaining 80% of the dataset were predicted with Q^2 = 0.998.