Shot Noise and Dose Optimization in Electron Beam Lithography: Balancing Roughness, Resolution and Write Time

Where shot noise comes from and how it scales

An electron beam lithography system does not deliver a continuous dose; it delivers electrons, and the number of electrons landing in a given exposed region during its dwell time fluctuates statistically even for a perfectly stable beam. The arrivals follow Poisson statistics: if the mean count is N, the standard deviation is √N, and the relative fluctuation is 1/√N. Write with 10,000 electrons in a pixel and the dose carries roughly one percent relative noise; write with 625 and the noise is four percent. Halving the dose quadruples the relative variance—this steep scaling is why shot noise bites exactly where users most want relief, at low doses on fine features.

The resist converts dose fluctuation into geometry fluctuation. Around the clearing threshold, the dissolution rate depends steeply on dose, so a few percent of dose noise becomes edge-position noise on every feature. Because the fluctuations are random and spatially uncorrelated, they appear as linewidth roughness and edge-placement randomness that resist chemistry, beam stability and staging precision cannot remove. Backscattered secondary effects add a correlated component, but the Poisson floor is the part set purely by counting statistics.

Finding the dose window with a matrix

The dose matrix is the standard instrument for the trade. Write the same test pattern—a layout containing the critical feature sizes, pitches and structures of the real work—across a range of doses spanning roughly a factor of two to four in logarithmic steps, holding resist stack, development and everything else fixed. Measure CD and LWR on each die with a stated method and enough samples per dose to see the trend. The result is a window: below it, features thin, roughen and vanish; above it, CDs inflate, dense patterns bridge through proximity, and the resolution degrades. The window's width is the dose latitude of the process.

Choose the working dose inside the window with margin from both edges—not at the clearing threshold and not at the first noisy point. The margin absorbs the variables a test matrix holds constant: resist batch variation, development drift, substrate backscatter differences between the test layout and production layouts. Two to five percent dose margin at each edge is a common working rule; tighter margins demand tighter process control, which is itself a cost.

• Matrix must use the real resist stack, substrate and development recipe
• Test pattern must contain the critical features, not a generic grid
• Measure LWR as well as CD—CD alone hides the noise floor
• Record sample counts; roughness measured on three lines is noise about noise

Dose, write time and the economics of noise

Write time scales with the total charge delivered: doubling dose doubles dwell times and doubles machine hours for a fixed beam current. This is the economic face of shot noise. On a sparse test layout the difference between two candidate doses is minutes; on a dense production layout with millions of shapes it can be hours per wafer. The optimized working dose is therefore not the one that minimizes roughness in absolute terms—it is the highest noise the pattern's LWR and CD targets tolerate, because that yields the lowest acceptable dose and the shortest schedule.

The trade interacts with machine capability. A system with higher available beam current at the same spot size and resolution delivers the same electron count in less dwell time, which is one of the practical meanings of throughput in EBL specifications. But current cannot be raised indefinitely: at high beam current the spot grows and resolution suffers, so the dose–time–resolution triangle is closed at every point. The honest way through is per-layout: compute write time at the candidate doses on the real data, and let the roughness targets decide how much time the noise floor is worth.

Proximity correction and the final working dose

Shot noise sets the floor, but the dose actually assigned to each shape is modulated by proximity effect correction: dense regions need less dose per shape because neighboring exposures add deposited energy, isolated shapes need more. PEC therefore multiplies the base dose by pattern-dependent factors, and both the noise and the process window travel with the modulated dose. A working dose chosen from a matrix run without PEC will not transfer exactly to corrected production patterns—the right procedure measures the final window with the production PEC settings applied to the test layout, or at minimum verifies CD and roughness on the first corrected production write.

Charging effects on insulating substrates shift local deposited energy the same way, and temperature-sensitive resists drift between matrix and production if development differs by even a minute. The lesson generalizes: dose optimization is valid only for the exact stack and conditions it was measured on, and every later change—resist batch, developer, substrate, beam energy—reopens the question. Treat the dose matrix as a living artifact, re-run when inputs change, and archive the curves with the recipe so the working dose always has its evidence.