Artificial Intelligence, Unified Machine Learning

Attractor Assisted Artificial Intelligence with Unified Machine Learning

In 1991, Solyman Ashrafi began a project at NASA GSFC. The purpose of this project was to identify ways to predict the re-entry footprint of Low-Earth Orbit (LEO) science satellites. It was known at the time that something similar happens to the Earth's atmosphere as a result of sunspots and sunspots are correlated with solar activity. As solar activity (measured at 10.7 cm radio wavelength) increases, extreme ultraviolet radiation (EUV) heats our planet's gaseous envelope, causing it to swell and reach farther into space than normal. While puffed-up, this atmosphere will increase drag on LEO satellites and drag them down in an orbit decay.

​

There are multiple layers in the atmosphere that extend 100s of km above earth. The first layer above the surface of the earth is the troposphere (10 km) where weather occurs. The second is the stratosphere (10 – 50 km) where many aircraft fly because it is much more stable. The third layer is the mesosphere (50 - 100 km) where meteors burn up. The fourth layer is the thermosphere (100 – 500 km). Aurorae occur in the lower thermosphere (100 – 150 km). The space shuttle used to orbit in the thermosphere as does the International Space Station (ISS). Orbit decay happens as the LEO satellites move through the thermosphere. The most famous example of this effect was Skylab, which burned up in the atmosphere after orbit decay.

​

Some satellites have onboard jets to compensate for orbit decay. Other LEO satellites need a little help. The Hubble Space Telescope (HST) has no jets, so the only way to change the altitude is to grab it and move it. This has been done by the space shuttle during service missions. Astronauts have also raised the altitude of the ISS in the past. However, some parts of the satellites can survive atmospheric re-entry and fall to Earth. Because of the increased number of satellites being launched, the flight dynamics group at NASA GSFC had a specific program for this issue.

​

NASA implemented guidelines to develop controlled and semi-controlled re-entry for satellites. Controlled re-entry is the ability to force the entry over a pre-determined area or region, where parts could fall in specific low-populated regions or areas that would allow them to capture the satellies with a net in order to parachute them to a ship to save costs or protect secrets. Because of the complex and changing interactions between the atmospheric density and solar activity, it was difficult to predict the footprint of the satellites unless we could model the solar activity. We had multiple models of the sun from thermonuclear, to electromagnetic, to mechanical models, to fluid/gas model, magnetic tubes of sunspots…etc.

​

Unfortunately, none of the models could be reconciled for predicting the radio solar flux 10.7 and therefore Dr. Ashrafi was left with trying to predict solar activity using data (collected daily by NOAA since the invention of radar in World War II). Therefore, this was a big data analysis before the word “big data” was even coined. That is, Dr. Ashrafi wanted to bypass the complicated modeling of the physics of the sun and just rely on how to predict the solar flux 10.7 entirely from a one dimensional time series, constructing the drivers of its dynamics and predicting its behavior without relying on the underlying physics. Many scientists thought that this would not be possible and when Dr. Ashrafi ultimately did it, it was recognized, and he received an achievement award.

​

Dr. Ashrafi started his data analytics by reconstructing the dynamics in a multi-dimensional phase space. This was done by using a topological technique of delay embedding. The reconstructed phase space of the dynamics is called an “attractor” because its geometrical shape attracts the trajectory of the dynamics in phase space. Once constructed, the number of embedding dimensions is the number of causal drivers of the dynamics (not correlative but causal drivers). This attractor also has some topological features and invariants that do not change and are therefore the characteristic of the dynamics. There are sophisticated techniques to extract these topological features and invariants. Once they are extracted, they can be used to predict or forecast future behavior without the knowledge of the underlying physics behind the phenomenon.

Dr. Ashrafi also created a formula that allows for the prediction of the dynamics up to a predictability horizon as a function of length and frequency of the past data, global and local Lyapunov exponents and the fractal dimension of the attractor. This allowed Dr. Ashrafi to predict solar activity better than statistical methods up to that horizon beyond which his predictions would approach the statistical predictions. What was remarkable is that Dr. Ashrafi could specify how much historical data was needed to achieve that horizon.

​

Two of the most critical items in development of an AI engine are addressed by our patents. One is the training data needed to model the system which could be specified based on our formula. That means if one would add more training data, there is a diminishing return on the accuracy or extension of the prediction. The second is that the embedding dimensions represent causal drivers of the dynamics. If we project the attractor on to each of these dimensions, we may be able to identify what those drivers are and therefore model the system. Today’s use of neural nets for forecasting includes several hidden layers and numbers of neurons within each of the layers with specific activation functions. However, it is not possible to deduce the causal relationships to build a mathematical or physical model of the system, but Dr. Ashrafi’s approach is capable of such modeling.

​

Therefore, this patent is called “Attractor Assisted AI” so that we do not blindly dump a lot of data on to the AI engine which helps control the speed and accuracy of predictions up to a specified horizon.