1. 背景与原文回顾
我们近期拟投稿的一篇英文论文,题为 “Data-driven modeling of cycle-averaged and phase-resolved aerodynamic responses of a three-dimensional rigid flapping wing”。初稿完成后,我让学生将论文交给 AI 助手做语言润色,结果 AI 只是机械地替换了几个同义词,完全没有触及根本问题——摘要写成了参数研究的流水账,Highlights 也与正文贡献脱节。
1.1 初稿摘要(原文)
The aerodynamic response of a flapping wing depends jointly on its operating condition and on the phase within the flapping cycle, posing distinct requirements for interpretable mean-load relations and phase-resolved surrogate models. Here, a data-driven analysis was conducted of a three-dimensional rigid NACA 0014 wing using 72 computational-fluid-dynamics (CFD) conditions spanning freestream velocity, flapping frequency, flapping-angle amplitude, and static angle of attack. Separate reference areas and a propulsion-positive sign convention were used to distinguish dimensional forces from normalized lift and thrust coefficients. Along the sampled one-factor sequences, increasing freestream velocity reduced both mean coefficients while increasing the corresponding forces; increasing frequency strengthened the aerodynamic responses overall; increasing amplitude primarily enhanced thrust without a monotonic lift benefit; and increasing angle of attack raised lift while inducing a thrust-to-drag transition between 7.5° and 10°. For cycle-averaged lift, a frozen symbolic-regression expression achieved a test RMSE of 0.0292 and $R^2=0.9940$, while a second-order polynomial Ridge baseline attained an RMSE of 0.0234. For phase-resolved prediction, a five-network deep-neural-network ensemble achieved range-normalized RMSE values of 1.80% for thrust and 1.77% for lift, compared with 7.66% and 7.07% for a Fourier–Ridge baseline. Errors increased for joint parameter changes and frequency extrapolation. These results establish complementary data-driven routes for explicit mean-load estimation and accurate phase-resolved waveform prediction across the sampled operating conditions.
1.2 初稿 Highlights(原文)
- Mean forces and coefficients exhibit distinct responses to freestream velocity.
- Symbolic regression provides a compact explicit relation for cycle-averaged lift.
- A DNN predicts phase-resolved loads accurately within explicit validity boundaries.
2. 核心问题诊断
2.1 摘要:参数趋势喧宾夺主
- 结构失衡:全文 8 句话中,第 4 句用分号并列了四个单因素趋势(速度、频率、幅度、攻角),占全文近 40% 篇幅。读者读完一半还不知道文章到底做了什么方法创新。
- 细节过度:攻角趋势中给出了精确角度范围 “between 7.5° and 10°”,这是正文结果部分的离散数据点,放在摘要中既破坏与其他趋势的详略平衡,又让读者误以为这是普适规律。
- 语言翻译腔:大量被动语态(“was conducted”, “were used”)、中式直译(“Along the sampled…”, “establish complementary data-driven routes”)、以及生硬短语(“without a monotonic lift benefit”)。
2.2 Highlights:与摘要和正文均不匹配
- 亮点 1 只提了速度对力/系数的不同影响,这是最基础的参数分析,并非文章核心贡献。
- 亮点 2 忽略了与 Ridge 基线的对比,未提及 SR 表达式的测试误差(RMSE 0.0292),信息量不足。
- 亮点 3 中 “within explicit validity boundaries” 表述不准确——文中并未给出严格边界,而是通过测试集分类展示了误差递增趋势;且未提及 DNN 的集成策略、对比基线及具体精度。
3. 修改方案:全新摘要与 Highlights
3.1 新摘要(修改版)
This study presents a data-driven framework for predicting cycle-averaged and phase-resolved aerodynamic responses of a three-dimensional rigid flapping wing. Using 72 computational-fluid-dynamics conditions spanning freestream velocity, flapping frequency, amplitude, and angle of attack, we develop two complementary models: a symbolic-regression expression for mean lift and a five-network deep-neural-network ensemble for phase-dependent thrust and lift. For cycle-averaged lift, the selected explicit expression achieves a test RMSE of 0.0292 and $R^2=0.9940$, while a polynomial Ridge baseline yields a lower RMSE of 0.0234, illustrating a trade-off between interpretability and accuracy. For phase-resolved prediction, the DNN ensemble attains range-normalized RMSE values of 1.80% for thrust and 1.77% for lift, outperforming a Fourier–Ridge baseline (7.66% and 7.07%). Errors are lowest for one-factor interpolation and increase systematically for joint parameter variations and frequency extrapolation, thereby defining empirical validity boundaries. These results establish a reproducible basis for selecting either an explicit mean-load relation or a phase-resolved surrogate in flapping-wing aerodynamic analysis.
3.2 新 Highlights(修改版)
- A data-driven framework couples symbolic regression (explicit) and a DNN ensemble (accurate) for flapping-wing aerodynamics.
- Symbolic regression yields a compact mean-lift expression (RMSE 0.0292), trading interpretability against a more accurate Ridge baseline (RMSE 0.0234).
- The DNN ensemble achieves approximately 2% range-normalized RMSE for phase-resolved thrust and lift, with errors increasing from interpolation to frequency extrapolation.
4. 修改思路详解:从“流水账”到“贡献导向”
4.1 结构重组(核心转变)
| 初稿结构(问题) | 新摘要结构(解决) |
|---|---|
| 第1句:背景(依赖条件和相位) | 第1句:直接点明“本文提出了一个数据驱动框架” |
| 第2句:数据细节 + 被动语态 | 第2句:数据规模 + 两个互补模型(核心贡献) |
| 第3句:符号约定(琐碎) | 合并到方法描述中,不单独成句 |
| 第4句:四个趋势流水账(占40%) | 完全删除,压缩为引言/正文内容 |
| 第5-6句:SR 和 DNN 结果 | 第3-4句:SR 和 DNN 结果(保留关键数字) |
| 第7句:误差增加(模糊) | 第5句:误差随测试类别的变化(明确边界) |
| 第8句:总结(routes across conditions) | 第6句:结论(reproducible basis for selecting models) |
为什么这样改?
英文期刊摘要更倾向“倒金字塔”:先亮核心贡献,再给关键证据,最后落脚到意义。初稿是典型的“循序渐进”中式写法,编辑和审稿人需要读到一半才能抓住重点。新摘要将参数趋势全部压缩为第2句的 “spanning freestream velocity…”,一笔带过,把节省的篇幅让给方法创新和性能对比。
4.2 语言去翻译腔
- 主动语态:全篇使用
We develop,we establish等,避免was conducted等被动。 - 动词精准:
presents(框架),develop(模型),attains(精度),outperforming(对比),defining(边界)。 - 逻辑连接:用
while(对比 SR 和 Ridge),thereby(误差导致边界),either...or(选择意义) 等自然衔接。 - 删减冗余:去掉
Here,,Along the sampled...,creating distinct requirements等中式直译。
4.3 Highlights 的同步调整
Highlights 必须与摘要的“贡献导向”一致,且每条应是一个独立、自包含的卖点。
- 第1条:删除旧亮点1(速度趋势),改为突出框架双模型结构,点明“显式 vs 准确”的互补性。
- 第2条:保留 SR 的 RMSE 0.0292,但补入 Ridge 的 0.0234 作为对比,明确“可解释性-准确性权衡”。
- 第3条:将具体数字 1.77–1.80% 改为 approximately 2%(只告诉读者大致误差量级),并将误差递增规律(插值→外推)作为边界依据,完全剔除不准确的“explicit validity boundaries”说法。
4.4 细节处理原则
- 具体数字:摘要中保留 SR 和 DNN 的精确 RMSE(因为这是方法性能的核心证据),但 Highlights 中 DNN 精度改用约数,符合“亮点”的概括性要求。
- 攻角推力-阻力转变:完全删除 7.5°–10° 这个具体区间,因为这不是文章贡献,且数字过于具体。正文 Results 部分已有完整数据,摘要无需重复。
- 符号回归的 “frozen”:改为
selected,更准确。
5. 给学生们的三点写作建议
-
摘要不是“实验报告”,而是“贡献宣言”
永远先问自己:如果读者只记住一句话,我希望是哪句?把这句话放在摘要最前面。参数影响、数据细节、符号约定都是支撑材料,能压缩就压缩,能删除就删除。 -
Highlights 是“广告语”,不是“目录”
每条 Highlight 应该让非专业编辑一眼看懂你的创新点,并愿意点开全文。避免模糊表述(如 “explicit validity boundaries”),用具体对比(如 “trading interpretability against accuracy”)和关键数字(但非全部细节)来支撑。 -
警惕“翻译腔”的三种典型症状
- 过度被动语态(尤其 “was conducted/were used”)
- 直译中文连接词(“Along…”, “Based on… we found that”)
- 抽象名词堆砌(“establish complementary data-driven routes”)
修改方法:写完摘要后,大声朗读,如果觉得拗口,就拆成短句或换主动动词。
6. 结语
这篇博客记录了从初稿到终稿的完整修改轨迹。希望同学们在下次写摘要时,能跳出“先介绍背景,再讲做了什么,最后说结果”的惯性,尝试用“倒金字塔”结构,把最硬的成果亮在开头。记住:编辑和审稿人没有义务读完你的摘要去猜你的贡献,你必须替他们总结好。
如有任何问题,欢迎在组会中讨论。