Following last year’s advancement in the Laboratory for Microsystems Technology –– creating an ultra-open metamaterial that silences a broader range of noise while preserving ventilation –– Professor Xin Zhang (ME, ECE, BME, MSE) and team have further improved upon the acoustic silencer, addressing key problems and advancing the utility of their instruments. A new paper has been published in Advanced Science, titled "Physics-aware machine-learning-driven inverse design of broadband ultra-open acoustic metamaterials."
    In what Zhang describes as "the next step in a longer progression" of enhancements, this year’s improvement focuses on a physics-aware, inverse-design framework: working backward from the ideal performance enhancement toward the design requirements needed to make it a reality. Through machine-learning algorithms and techniques, the lab is able to cover a broad range of potential performance enhancements across a much shorter timeframe.

In essence, the latest improvement is not only on the acoustic silencer, but the means by which to identify and design adaptations for a wider diversity of needs and applications. "The key advance [is not] simply a better silencer, but a new way to design silencers for different needs," Zhang explains. "The resulting designs can silence more than 830 Hz of sound bandwidth while remaining up to 80 percent open to airflow and only 4–8 centimeters thick."
    The improvements don’t stop there, however, as the team has also focused on advancing scattering-based phase physics through machine learning techniques.
    Scattering-based phase physics, in the context of the acoustic silencer technology, operates like this, according to Zhang’s team: sound is traveling through two unique pathways within the instrument, and while one moves directly, the other is intentionally delayed in order to cancel out the initial wave. Doing so reduces any outgoing sound while preserving a highly open path for airflow.
    How does machine learning play into this improvement? Once more through the backward-step process, machine learning allows the team to interface with a wide variety of design variations in order to find the most effective "geometric relationships" that produce optimal acoustic interference. Just as importantly, the framework reveals previously hidden design rules linking geometric relationships to acoustic performance, helping the researchers understand not only the "how" behind the designs, but the "why."

With over 200,000 design-response examples, this machine learning framework fits neatly into the timeline of the Zhang lab’s overall project: creating open structures to silence sound across a wide variety of needs and applications. While last year’s advancement proved the broader applicability of the acoustic silencer, this year carves a path toward more specific designs and methods for utilizing the lab’s silencing technology. In other words, accelerating design and increasing the discovery of new applications.
    This is not, it should be noted, a one-size-fits-all approach; rather, Zhang’s team is aiming to offer a customization framework with their instrument.
    By improving machine learning capabilities, making it more cognizant of design needs and constraints––hence, physics-aware––engineers can craft and pick a design that, as Zhang says, "best satisfies those additional requirements."

So what’s next for the Laboratory for Microsystems Technology, and the acoustic silencer? According to Zhang, the short-term goal is to pioneer even greater application-specific designs, further utilizing requirements such as "noise spectra, airflow, available space, and manufacturing constraints." Moreover, they want to address areas for improvement in their current design applications by tackling greater attenuation, ventilation, and compactness needs.
    "Longer term, [our] goal is a design-to-device workflow," Zhang says, "in which an engineer specifies the desired acoustic performance and the framework proposes physically realizable structures suited to that application."

Importantly, the Zhang lab emphasizes that this is not an approach through which AI is replacing the utility of physics, but achieving a sort of symbiosis in order for each to improve the other. In their own words: "the goal is not to replace physics with AI, but to combine them so that each makes the other more powerful."