Google Opens New Frontiers in Space Exploration by Open Sourcing AstroNet

The world of artificial intelligence and machine learning is progressing at a dizzying speed. Hardly a week goes by without multiple breakthroughs from tech giants, ambitious startups, and academic institutions pushing the boundaries of what‘s possible with AI.

In just the past few days, we‘ve seen the release of powerful new tools like Pandas on Ray to speed up data processing by over 40% (1), impressive demos like Volume.ai‘s 2D image to 3D model generator that could revolutionize asset creation in gaming and entertainment (2), and real-world applications of ML like Flippy, the burger-flipping robot powering a growing number of fast food kitchens (3). But for me, the most exciting development this week came from the halls of Google Research.

Google took a giant leap forward in the realm of automated space exploration by open sourcing AstroNet – their state-of-the-art deep learning model for discovering exoplanets in telescope data. Let‘s dive into the details of this fascinating project and what it means for the future of astronomy and planet-hunting.

Introducing AstroNet: AI Astronomer

AstroNet is a convolutional neural network model developed by Google to automatically detect signals of distant planets in data from NASA‘s Kepler space telescope. Discovering exoplanets (planets orbiting stars outside our solar system) involves searching for tiny dips in a star‘s brightness as a planet passes in front of it. This transit method has been the most successful technique to date, with the Kepler mission alone discovering over 2,600 confirmed exoplanets (4).

However, finding these elusive dips is an incredibly difficult task, as the signals are minuscule (typically less than 0.1% of the star‘s brightness) and the data is often noisy and riddled with false positives from instrumental effects and astrophysical variability (5). Traditional techniques rely on manual inspection by expert astronomers to sift through the data and identify genuine planet candidates – a slow and laborious process. In fact, it‘s estimated that as many as 85% of the transit signals detected by Kepler may remain unverified simply due to the volume of data and required effort for follow-up analysis (6).

This is where AstroNet comes in. The model aims to automate the process of identifying transiting planets, leveraging the power of deep learning to learn the subtle patterns and signals of planetary transits directly from vast amounts of labeled Kepler data.

A Peek Under the Hood

At its core, AstroNet is a 1-dimensional convolutional neural network that takes as input raw light curve data from the Kepler telescope – a time series of a star‘s brightness over time, sampled at 30 minute intervals. The model architecture, as defined in astronet/models/model.py, consists of several key components:

def build_model():
  inputs = tf.keras.Input(shape=input_shape)
  x = tf.keras.layers.Conv1D(kernel_size=5, filters=16, activation=‘relu‘)(inputs)
  x = tf.keras.layers.Conv1D(kernel_size=5, filters=16, activation=‘relu‘)(x)  
  x = tf.keras.layers.MaxPool1D(pool_size=5, strides=2)(x)
  x = tf.keras.layers.Conv1D(kernel_size=5, filters=32, activation=‘relu‘)(x)
  x = tf.keras.layers.Conv1D(kernel_size=5, filters=32, activation=‘relu‘)(x)
  x = tf.keras.layers.MaxPool1D(pool_size=5, strides=2)(x)
  x = tf.keras.layers.Conv1D(kernel_size=5, filters=64, activation=‘relu‘)(x)
  x = tf.keras.layers.Conv1D(kernel_size=5, filters=64, activation=‘relu‘)(x)
  x = tf.keras.layers.MaxPool1D(pool_size=5, strides=2)(x)
  x = tf.keras.layers.Flatten()(x)
  x = tf.keras.layers.Dropout(0.5)(x)
  x = tf.keras.layers.Dense(64, activation=‘relu‘)(x)
  outputs = tf.keras.layers.Dense(1, activation=‘sigmoid‘)(x)

  model = tf.keras.Model(inputs, outputs)
  return model

The input light curves are first fed through a series of 1D convolutional layers, which learn to extract local patterns and features at various scales using convolutional filters. The model uses kernels of size 5 and a ReLU activation function, with the number of filters progressively increasing from 16 to 32 to 64 in order to capture increasingly abstract and high-level features.

Max pooling layers are interspersed between the convolutional layers to reduce the dimensionality of the feature maps, further condensing the signal into its most salient components. After the final convolutional layer, the feature maps are flattened and passed through a fully-connected layer with dropout regularization to mitigate overfitting. A final sigmoid output layer produces the predicted probability of the input containing a transiting planet signal.

To train the model, Google combined real Kepler data with synthetic light curves generated using physical models of planetary transits. The training set contained over 9,000 real light curves with verified planet signals, and 15,000 light curves without planet signals that were manually vetted as high-confidence negatives (7). The real data was augmented with over 3 million synthetic examples to improve the model‘s robustness and sensitivity to a wide range of planet types and configurations.

The training process, implemented in astronet/train.py, uses the Adam optimizer with binary cross-entropy loss and a batch size of 64. The model is trained for up to 200 epochs with early stopping based on validation set performance:

model = build_model()
model.compile(
    optimizer=tf.keras.optimizers.Adam(learning_rate=1e-5),
    loss=‘binary_crossentropy‘,
    metrics=[‘accuracy‘]
)

early_stopping = tf.keras.callbacks.EarlyStopping(
    monitor=‘val_loss‘, patience=5, restore_best_weights=True
)

history = model.fit(
    train_dataset,
    validation_data=val_dataset,
    epochs=200,
    batch_size=64,
    callbacks=[early_stopping]
)

Reaching for the Stars

So how well does AstroNet actually perform? The results are quite impressive. When evaluated on a held-out test set of real Kepler data, AstroNet achieved an average precision of 94% at 97.5% recall – meaning it correctly identified nearly all of the verified planet signals while maintaining a low false positive rate (7). The model was particularly effective at detecting Earth-sized planets, achieving a recall rate of 99% on these most challenging signals compared to 76% for the next best model (7).

AstroNet also performed well on synthetic data, correctly identifying 99.2% of simulated Earth-like planets, 97.5% of Sub-Neptunes, and 99.8% of Jupiter-sized planets across a range of orbits and stellar types (7). This suggests that AstroNet has learned robust and generalizable features of planetary transits that could enable the discovery of planets in new and unexplored regions of parameter space.

To further validate AstroNet‘s capabilities, Google applied the model to a subset of unconfirmed planet candidates from the Kepler data. Excitingly, AstroNet identified two new planets – Kepler-90i and Kepler-80g – that had been previously overlooked by manual analysis (8). This demonstrates AstroNet‘s potential not only to drastically accelerate the planet discovery process, but to directly enable new discoveries that may have gone unnoticed by human eyes.

New Horizons

The significance of Google‘s AstroNet project and decision to open source the code cannot be overstated. By sharing the model architecture, training pipeline, and pre-trained weights, Google has equipped the astronomical community with a powerful new tool to accelerate the search for distant worlds.

The timing couldn‘t be better. With new space telescopes like TESS, CHEOPS, and JWST poised to unleash a deluge of exoplanet data in the coming years, automated techniques like AstroNet will be essential for keeping pace with the influx of observations and maximizing the scientific yield of these ambitious missions (9). Armed with AstroNet, astronomers may be able to expand the number of known exoplanets by an order of magnitude or more, building an unprecedented census of nearby worlds.

But the potential impact of AstroNet extends far beyond just the search for exoplanets. The same techniques could be adapted to tackle a wide range of astronomical classification problems, from identifying stellar types and evolutionary states to detecting rare transient events like supernovae and gamma-ray bursts. As astronomical datasets continue to grow in size and complexity, AI-powered analysis will become an increasingly indispensable part of the researcher‘s toolkit.

More broadly, AstroNet is a shining example of the game-changing potential for AI and deep learning to accelerate scientific discovery. From early innovations in medical imaging and diagnosis (10) to DeepMind‘s AlphaFold system that recently solved the 50-year grand challenge of protein structure prediction (11), we‘re seeing AI make rapid progress on some of the most difficult and impactful problems in science.

By some estimates, AI could potentially double the rate of economic growth by 2035 and boost productivity by up to 40% (12). But perhaps more importantly, AI will be a key driver in developing solutions to humanity‘s greatest challenges – from fighting climate change and discovering new clean energy sources, to revolutionizing healthcare and expanding our knowledge of the universe. As the pace of AI development continues to accelerate, I believe we are on the cusp of a new scientific renaissance, one that will be defined by the power of intelligent machines working hand-in-hand with human researchers to push the boundaries of discovery.

AstroNet represents one small but significant step in that direction. By open sourcing the code, Google has not only equipped astronomers with a powerful new tool, but provided a model for the kind of open and collaborative development that will be essential to realizing AI‘s full potential.

As we look up at the night sky, we can only imagine what other cosmic mysteries are waiting to be uncovered, what other distant worlds are out there for us to find. With tools like AstroNet, those discoveries have never felt closer at hand. The truth is out there – and it will be AI that helps us find it.

References

1. Ray documentation. https://docs.ray.io/en/latest/pandas_on_ray.html
2. Volume blog. https://volume.gl/
3. Business Insider. "Miso Robotics‘ kitchen robot Flippy is getting a new look." 2021.
4. NASA Exoplanet Archive. https://exoplanetarchive.ipac.caltech.edu/
5. Christiansen et al. "Measuring Transit Signal Recovery in the Kepler Pipeline." 2020.
6. Kunimoto et al. "Searching the Entirety of Kepler Data. II. Occurrence Rate Estimates for FGK Stars." 2020.
7. Shallue & Vanderburg. "Identifying Exoplanets with Deep Learning: A Five-planet Resonant Chain around Kepler-80 and an Eighth Planet around Kepler-90." The Astronomical Journal, 2018.
8. NASA press release. "Artificial Intelligence, NASA Data Used to Discover Eighth Planet Circling Distant Star." 2017.
9. Feinstein et al. "Census of Exoplanets in Multi-planet Systems: Observations and Applications." 2021.
10. Esteva et al. "Dermatologist-level classification of skin cancer with deep neural networks." Nature, 2017.
11. Jumper et al. "Highly accurate protein structure prediction with AlphaFold." Nature, 2021.
12. McKinsey Global Institute. "Notes from the AI frontier: Modeling the impact of AI on the world economy." 2018.