Phones and computers have gotten exponentially smarter and faster over the years. In the world of electronics, this growth has been made possible by Moore’s Law: the number of transistors on circuit boards that make up the heart of electronics doubles approximately every two years.
Nevertheless, the sector is reaching a juncture where integrating additional transistors onto silicon wafers, the traditional material for circuit boards, is proving to be a formidable challenge.
Consequently, materials scientists are investigating new materials and technologies to sustain the progression of computational power without being constrained by the physical limitations of silicon.
Principal Scientist Alagappan Gandhi, and Director for the Electronics and Photonics Department, Jason Png, from A*STAR’s Institute of High-Performance Computing (IHPC) have been exploring a novel paradigm that harnesses the power of light. “In contrast to electronic chips reliant on electrons, photonic chips exploit light for data transmission, yielding swifter communication and higher capacity,” noted the duo.
This photonic chip technology is similar to optical fibres used in long-distance data transfer, and according to experts, can completely revolutionise computer architecture.
While the potential of photonic circuits is vast, current chips have a much larger physical footprint compared to electronic circuits, presenting a hurdle for commercial viability. Addressing this, Alagappan and Png have reconceptualised the design of pivotal photonic components. Their study centred on the refinement of power dividers, integral devices that split light into discrete pathways for downstream processing.
Employing deep learning, the team enhanced inverse design algorithms—computational methods that ideate components conforming to specified functionalities and constraints.
Listing the merits of this approach, Alagappan said, “Embracing inverse design methodologies opens the door to creating innovative power dividers characterised by compactness, minimal loss, large bandwidth and importantly, suitability for large-scale manufacturing processes.”
Deep-learning models are advanced computer programmes that enhance the design process (which traditionally could be quite slow) by simulating how light moves through different structures. This allows for a more efficient and creative approach to design, where the computer can quickly test and refine different models.
This method proved to be particularly useful as the team successfully designed a range of power dividers that were shown to be highly efficient, compact and resilient to manufacturing variations.
Significantly, the team's deep-learning-based design process surpasses traditional methods in speed without sacrificing the precision or effectiveness of the final product.
Looking ahead, the researchers intend to apply their methods to other photonic components, potentially unlocking new prospects in quantum communications and optical networking.
The A*STAR-affiliated researchers contributing to this research are from the Institute of High-Performance Computing (IHPC).
