Yang Su retweeted
"a seesaw but for em waves" @suyangcn on the beauty of the patch antenna!
Today, we’re excited to unveil Heaviside-1, @arenaphysica's second generation foundation model for electromagnetism. 5 months since the release of Heaviside-0, Heaviside-1 is a crucial milestone on our quest to build EM superintelligence: a foundation model that understands EM across the spectrum, from RF through photonics, capable of designing the next generation of electronics.
Major updates:
- Heaviside-1 is >10x the size of Heaviside-0 (roughly the size of GPT-2), trained on 250k unique designs with >500B unique EM field samples.
- It runs 10^5 x faster than commercial solvers, with accuracy <1 dB.
- Heaviside-1 natively encodes 3D structures, their material properties, and excitation patterns as input.
It predicts full EM fields at arbitrary locations in space, not just downstream quantities like S-parameters.
- Most importantly, our bet on fields lets Heaviside generalize OOD proving that it’s learning real EM physics: Heaviside-1 generalizes outside of distribution very well, leaping from 0.99 dB to 0.53 dB S-parameter error.
Try it out in Atlas Fields Studio in beta today (link in replies). I personally had a lot of fun seeing how EM fields twist and curl around different circuits. I wish I had this when I was learning EM in college. I hope you enjoy it as much as we did.
Yang Su retweeted
Today, we’re proud to unveil Heaviside-1, @arenaphysica's second generation foundation model for electromagnetism.
Five months since announcing Heaviside-0, today’s launch marks a critical milestone towards building electromagnetic superintelligence, a foundational intelligence that understands EM across the spectrum and is capable of designing next gen electromagnetically-governed systems, from RF to semiconductors to photonics.
Why we’re so excited about Heaviside-1, and what today’s announcement represents:
> It’s big: Heaviside-1 is a GPT-2-size model, roughly 10x the size of Heaviside-0 and trained on 250k unique designs with >500B unique EM field samples
> It’s fast: It generates predictions with 5 orders of magnitude acceleration relative to commercial solvers, with under 1dB of accuracy
> Full-3D Encoding: Heaviside-1 natively encodes 3D structures, their material properties, and excitation patterns as input, a major step towards generalizability across design spaces
> Understands the underlying physics primitive: Heaviside-1 generates full-3D near-fields at arbitrary locations, not just downstream quantities like s-parameters
> Generalization: Most importantly, our bet on fields as the primitive for Heaviside-1 generalizes OOD, proving that it’s learning real EM physics. Heaviside-1 generalizes outside of distribution incredibly well and better than Heaviside-0, down to 0.53 dB s-parameter error from 0.99 dB on Heaviside-0.
Read more about Heaviside-1 and the work that went into it in this write-up by our research team, @m__frei, @chris_m_bryant, @nathanmirman, @HaoLiu69889370, Tommaso, Roberto, Ruichen, Noah, Boyuan, and @Trevs_Dev, and our RF team @hk2532_harish, @anrfic, @suyangcn and Zhaoji.
arenaphysica.com/publication…