One hundred meters underground in the Stringlight Research Institute, within the vast white space of the ultraclean room, only the continuous, low hum of the air purification system persisted, like the breath of a slumbering giant. The airflow here passed through layer upon layer of filtration, the temperature maintained within a fluctuation range of five-thousandths of a degree Celsius, and humidity rigidly held at the golden node of forty-five percent. Xiuxiu stood before the massive glass window of the observation gallery, clad in full-body cleanroom attire. Beneath the transparent hood, her gaze swept like the most precise probe over the colossal, complex apparatus below—a synthesis of human wisdom and the ultimate aesthetic of industrial craftsmanship—the "Stringlight One" High NA EUV lithography machine prototype.
It was no longer the "bonsai" of the laboratory, requiring careful nurturing and repeated parameter adjustments, but a "warhorse" poised to enter the production line and face the rigorous trials of industrialization. The Lithography Machine Division had officially been established the previous month, and the burden on Xiuxiu's shoulders shifted from conquering the principle-based "from zero to one" to the more arduous and tedious "from one to excellence"—transforming this prototype into an industrial product capable of stable, efficient, mass production.
The cheers celebrating the prototype's success seemed only yesterday, yet Xiuxiu and her team already clearly heard the ticking of a countdown. The market, competitors, and even expectations at the national level would not grant them much time to revel. The true battle had only just begun. And the core metric of this battle was no longer whether they could etch that slender, nearly abstract line, but a cold, objective number determining commercial success or failure, even the fate of a nation's industry—**throughput**.
Throughput: the number of wafers processed per unit of time (typically per hour). This is the ultimate manifestation of the commercial value of a lithography machine, especially an EUV lithography machine worth hundreds of millions of dollars, serving as the core bottleneck in chip manufacturing. It directly determines chip production costs, delivery cycles, and even the operational efficiency of the entire semiconductor industry chain.
Xiuxiu's fingertip gently traced the cold glass, as if she could feel the storm of speed and precision brewing inside the machine below. She took a deep breath; clean, dry air filled her lungs, bringing an almost cruel clarity. Improving throughput was an endless arms race centered on three core factors: **light source power, scanning speed, yield**. These three parameters were coupled and constrained, like a precise gyroscope—any imbalance in one could cause the entire system to collapse.
Her thoughts first focused on that pulsating, fierce "heart"—the **extreme ultraviolet light source**.
Extreme ultraviolet light at 13.5 nanometers—its generation itself resembled a miniature stellar birth. Inside the vacuum chamber, a high-power carbon dioxide laser, at a frequency of tens of thousands of times per second, precisely bombarded falling droplets of high-purity tin. Instantly, the tin droplet was heated, vaporized, ionized, forming a high-temperature plasma. In this brief, fleeting moment, the excited tin ions, upon returning to their ground state, would emit extreme ultraviolet light concentrated around 13.5 nanometers.
Yet this process was extremely inefficient—a luxurious energy conversion. Enormous laser energy input ultimately converted into usable 13.5-nanometer EUV light at less than a few percent. More energy was lost as heat, visible light, ultraviolet light, and high-speed splattering tin debris. This frustratingly low conversion efficiency was the first, most stubborn shackle on EUV light source power enhancement.
"Dr. Xiu, light source data," Xiuxiu said into the built-in communication microphone, her voice somewhat muffled within the hood.
The transparent AR interface before her immediately projected a real-time data stream. Light source power stabilized at 285 watts. This was a number the team could be proud of—during the prototype stage, they had lost sleep to break through the 250-watt commercial threshold. 285 watts meant that, under ideal conditions, it could support a certain scanning speed, achieving preliminary mass production capability.
But only preliminary.
"Plasma stability, CT3 level," came a young researcher's voice through the channel, carrying a barely perceptible tension. CT3 meant the plasma morphology exhibited minor fluctuations, not yet directly affecting exposure uniformity but lurking like a hidden time bomb.
Xiuxiu's gaze swept over the subtle, periodic jitters on the data chart. She knew the problem lay in tin droplet generation, requiring extremely high uniformity and timing. Liquid tin, under high pressure, passed through specially designed nozzles and was "cut" by high-frequency sound waves into tiny droplets merely tens of micrometers in diameter, flying toward the laser's focal point. Any minute disturbance—nozzle wear, jitter in the acoustic drive signal, even slight variations in tin purity—could cause deviations in droplet shape, speed, or trajectory. The laser bombarding an imperfect droplet naturally produced a "restless" plasma.
"Notify the materials group and fluid control group: conduct lifespan assessment on nozzle three and check the filtering parameters of the drive circuit," Xiuxiu's voice was calm yet carried unquestionable determination. "We need to raise stability to CT2 level—that's the baseline."
"Understood, Dr. Xiu."
Power and stability were like two ends of a balance. Blindly pursuing power enhancement might exacerbate plasma instability, leading to decreased exposure quality and ultimately lowering overall throughput. Being overly conservative, however, would fail to unleash the scanning system's potential. Xiuxiu had to find the optimal equilibrium point between them. She recalled Mozi occasionally mentioning "risk-reward ratios" in financial models; the corner of her mouth twitched slightly—the underlying logic seemed somewhat similar.
Having addressed the light source's "power" issue, next was guiding this faint yet precious EUV light on its journey of precise "carving." Once this beam left the plasma, it embarked on a march filled with "hostility."
First, **vacuum**. EUV light at 13.5-nanometer wavelength could be absorbed by almost any substance, including air. Therefore, from light source to objective lens to wafer stage, the entire optical path had to be maintained in a high-vacuum environment; any trace gas molecules would be "killers" of EUV light. Maintaining such a vast system's high vacuum was itself a colossal engineering challenge—the pumping efficiency of vacuum pumps, outgassing rates of sealing materials, even the time required to re-establish vacuum after each maintenance opening—all indirectly affected the machine's effective operating time and, by extension, throughput.
Second, **collection mirrors**. Because EUV light was so easily absorbed, traditional transmissive optical schemes completely failed; one had to employ all-reflective **Bragg reflectors**. These were not ordinary mirrors; they consisted of dozens, even hundreds, of alternating stacked silicon and molybdenum nanoscale thin films. The thickness of each film pair was controlled with extreme precision, precisely so that the incident EUV light, reflecting at each interface, would have its reflected waves superimpose and reinforce, achieving relatively high reflectivity. Yet even the most advanced multilayer structures had single-reflection reflectivity around only 70 percent. And the optical path required over ten reflecting mirrors; the light energy reaching the wafer from the source, after enduring countless hardships, amounted to less than two percent of the original.
This faint light, like a candle in the wind, was tasked with "writing" the blueprint of future chips onto photoresist. Therefore, any loss of light energy was unacceptable. Mirror contamination—whether deposition of tin debris or adsorption of residual hydrocarbons in the vacuum environment—would cause further reflectivity decline. Xiuxiu's team developed sophisticated **in-situ cleaning technology** for this, injecting trace amounts of hydrogen into the vacuum chamber. Under plasma action, hydrogen radicals generated would react with contaminants on the mirror surface, forming volatile substances then pumped away. This was like constantly performing non-contact cleaning on an invaluable antique mirror, requiring precise control of hydrogen flow rate, injection timing, and plasma parameters. The slightest misstep—ineffective cleaning was minor; damaging the precious multilayer structure would be catastrophic.
"Objective lens group thermal deformation data," Xiuxiu ordered again.
The AR interface switched, displaying the real-time temperature fields and micro-strain distribution of the aspheric mirrors forming the objective lens system, transmitted back by hundreds of embedded sensors. Even with such low reflectivity, the small portion of energy carried by EUV light absorbed by the mirrors converted into heat. Non-uniform thermal loads caused the mirror surfaces to undergo minuscule, nanometer-scale deformations—for EUV lithography pursuing atomic-level precision, this was enough to cause disastrous drops in imaging quality.
To solve this, the mirrors were embedded with complex microchannel cooling systems, circulating ultra-high-purity coolant with temperature control accurate to thousandths of a degree, to "suppress" thermal deformation. Additionally, the system integrated an **active deformation compensation system**. By real-time monitoring mirror surface deformations, it drove hundreds of **piezoelectric ceramic actuators** on the mirror backs, applying precise stress to "bend back" distortion caused by heat. This was like equipping the mirrors with a highly sensitive set of "muscles and nervous systems," enabling them to resist their own "fatigue deformation" from heating.
"Coolant loop B, zone three temperature fluctuation exceeds threshold by five percent," the system emitted a soft warning tone.
"Switch to backup circulation pump, check main pump drive unit," Xiuxiu responded almost without thinking. Such seemingly minor faults were commonplace in daily operations; any tiny slip in any link could cause the entire machine to go down, throughput directly zero. Reliability was the most fundamental guarantee of throughput.
When the light source provided sufficient power, stable output of EUV light, and the optical system could efficiently, precisely project the mask pattern onto the wafer, next was making the wafer "move" with the highest efficiency to receive exposure. This was the second key factor affecting throughput—**scanning speed**.
Lithography was not a static process. It required the precise "dance" of **dual-stage wafer tables**. One stage carried the wafer already aligned and awaiting exposure, moving beneath the objective lens for high-speed, stable scanning motion; meanwhile, the other stage carried the next wafer to be exposed, undergoing equally precise, nanometer-level alignment and leveling operations under the measurement system. When one wafer completed exposure, the two stages exchanged instantly, almost "seamlessly" beginning the next wafer's exposure.
This exchange process was called the "stepping" time within the "step-and-scan" cycle. Shortening stepping time meant reducing idle waiting, increasing the number of wafers processed per unit time. This placed extreme demands on the wafer stages' movement speed, positioning accuracy, and the degree of coordination in synchronization control between the two stages.
The wafer stage's motion was not simple linear acceleration and deceleration. It needed to accelerate from rest to several meters per second within milliseconds, then maintain extremely stable uniform motion (scanning) over the exposure area, finally decelerating precisely to zero. Throughout this process, any minuscule vibration, overshoot, or jitter would directly cause line blurring or overlay error exceeding standards.
"Initiate scanning synchronization test, sequence Alpha-7," Xiuxiu ordered. She needed to see with her own eyes whether this dual-stage system she had poured countless efforts into could maintain that effortless stability when pushed toward theoretical limits.
Through the observation window, she saw the two massive platforms below carrying simulated wafers, like elegant dancers on ice, begin accelerating. Air bearings kept them floating in an almost frictionless state; linear motors provided fierce yet precise power. Acceleration, stabilization, scanning, deceleration, exchange... movements flowed seamlessly; data cascaded waterfall-like on the AR interface, displaying real-time curves of position, velocity, acceleration.
"Y-axis stage: high-frequency jitter of 5 nanometers detected at the end of the scanning interval," reported a calm female voice—the head of the stage control team.
Xiuxiu's brow furrowed slightly. Five nanometers—for them already in the sub-nanometer control era, this was a number demanding serious attention. This high-frequency jitter might stem from structural resonance, slight poor adaptation of control algorithm parameters under extreme conditions, or merely minute fluctuations in bearing lubrication gas pressure.
"Record all sensor data, especially vibration sensors and current loop feedback. Notify the control algorithm group: 2 p.m., first meeting room for data consultation," Xiuxiu said gravely. "We need to find the root cause of the jitter—whether it's a mechanical structure issue or the control loop needs re-tuning."
Improving scanning speed was like tuning an F1 race car—every tiny parameter optimization might bring lap time improvement but could also push the system toward instability or even collapse. This was close-quarters combat with physical limits.
Yet merely having speed and power was far from enough. If most exposed wafers were scrapped due to defects, then even the highest throughput would be a castle in the air. This was the third factor affecting throughput, and the most elusive—**yield**.
Yield: the percentage of qualified chips relative to total production chips. In the EUV era, factors affecting yield were as numerous as hairs and extremely hidden. **Random defects** were among the most troublesome enemies. They might stem from a tiny, composition-inhomogeneous cluster in the photoresist; a nearly invisible, native defect on the mask from the manufacturing process; a nanometer-scale particle breaking through layers of filtration from the environment; or even the **random fluctuation** effect caused by the inherent randomness of EUV photon emission itself.
Especially in the era pursuing higher resolution, smaller linewidths, the probability of random defects increased exponentially. Like a painter using an increasingly fine brush—any tiny speck of dust falling on the canvas might ruin the entire work.
Xiuxiu switched to the defect detection system interface. The screen displayed an electron scanning image of a test wafer just completed exposure and development. Algorithms marked suspected defect locations with red circles—densely packed, over a hundred.
"Classification results," Xiuxiu's voice carried a hint of weariness.
"Sixty percent point defects, preliminarily judged related to photoresist composition fluctuations; thirty percent line edge roughness exceeding standards, possibly from combined effects of photon noise and colloidal properties; remaining ten percent, source unknown, requiring further cross-sectional analysis."
"Source unknown..." Xiuxiu murmured. These "source unknown" defects were the most troublesome. They appeared and disappeared like ghosts, unpredictable, un-eradicable, severely constraining yield improvement.
She thought of Yue'er. During the last three-way video call, when Xiuxiu complained about these elusive random defects, Yue'er had thoughtfully said, "Mathematically, this resembles a random point process in a high-dimensional space. Perhaps we could try using **stochastic geometry** and **point process statistics** tools to build probability models for these defects—not predicting each defect's appearance, but grasping the 'patterns' and 'clustering tendencies' of their occurrences."
At the time, Xiuxiu thought it was just a theoretical mathematician's beautiful daydream. But now, facing those glaring red circles on the screen, something stirred within her. Perhaps it really could be attempted? Inputting defect locations, morphologies, sizes, and other parameters into a model, analyzing their spatial distribution patterns on the wafer surface—whether hidden correlations existed with fluctuations of certain process parameters? At least this offered a new approach, a way to understand and combat randomness from probability and statistics perspectives.
"Package and encrypt this defect distribution data, along with corresponding process logs, send to Professor Yue'er's dedicated server," Xiuxiu instructed her assistant. "Include my access request and problem description."
Having dealt with yield challenges, Xiuxiu's thoughts drifted toward a more distant future. The existing "single-beam" scanning approach would eventually hit physical limits in throughput improvement. Light source power couldn't increase indefinitely; scanning speed had its ceiling; yield control grew increasingly difficult. To achieve the next leap, paradigm breakthroughs were necessary.
Her gaze turned toward the sealed apparatus in the laboratory corner—still in the proof-of-concept stage—an early prototype of **multi-beam parallel writing technology**.
Its core idea was like replacing a single, back-and-forth scanning "pen" with tens of thousands of "micro-pens" that could write simultaneously. Through a technology called **micro-projection array**, a single EUV beam would be split into tens or even hundreds of thousands of micro-beams, each independently controllable, corresponding to a tiny region on the chip. Thus, there would no longer be a need for wafer stages carrying entire wafers in large scanning motions; only tiny stepping would be needed, and huge areas could be exposed simultaneously, theoretically boosting throughput by several orders of magnitude.
Yet the technical difficulties therein were daunting. How to generate and control tens of thousands of EUV micro-beams? How to ensure absolute uniformity and synchronization among them? How to handle the massive data throughput and real-time control issues that ensued? How to design corresponding masks? Each question could occupy a top-tier R&D team for years, even decades.
Xiuxiu knew this might be a technological path considered for "Stringlight Two," even "Stringlight Three." But it was like a lighthouse in the distance, guiding direction, preventing her and her team from getting lost in the mire of details while dealing with the complex, intricate practical problems before them, always maintaining imagination for the future and the courage to push forward.
She once again looked through the observation window at "Stringlight One" below. It stood there quietly; beneath its cold metal shell flowed rushing photons, violent plasma, precise mechanical motion, and the complex symphony composed of countless lines of control code. Improving its throughput was a marathon without end, the ultimate optimization of the three core factors—power, speed, yield—a deep integration and challenge of interdisciplinary knowledge spanning materials, physics, chemistry, mechanics, control, software.
She felt a touch of weariness, but more, an excitement and calm immersed in immense challenges. That night she resolutely returned from the Netherlands, tears flowing alone in her apartment; the doubtful gazes of her team during the initial DUV battle; her standing her ground when conquering immersion technology; the darkest moments when EUV light source power couldn't break through; the excited embrace when dual-stage synchronization succeeded... scenes flashed through her mind.
This road, she had walked for ten years. And this "throughput" battle of extreme ultraviolet before her was merely another fortress that must be conquered on this long journey. She adjusted her breathing, and through the AR interface, began reviewing the next experimental report on improvements to the chemical amplification mechanism of photoresist.
The battle continued. The light, sleepless and unceasing.